Glossary
AI glossary
420 AI terms in plain words. Each links to the lessons that teach it.
#
- 95th percentile latency (also: p95, tail latency)
- The response time that 95% of requests beat. It shows what slow users experience, which an average hides.
- Learn it in: Latency budgets and load testing
- Related: Latency, Latency budget
A
- Acceptance criteria
- The specific, testable conditions a deliverable must meet for the client to accept it.
- Learn it in: Scoping an automation project, Writing a scope of work
- Related: Scope of work
- Accountability
- Clear responsibility for an AI system's actions and outcomes: who owns it, who approved its use, who answers when it causes harm, and the records that show what happened.
- Learn it in: Your company's AI policy, Designing human in the loop, Writing an AI policy, Accountability for agent actions
- Related: Audit log, Human in the loop, Responsible AI
- Adoption
- How many of the intended users actually use a system, how often, and for the tasks it was built for.
- Learn it in: Designing human in the loop, Change management and training, The KPI that matters
- Related: Change management, KPI
- Adversarial example (also: evasion attack, adversarial input)
- An input crafted to make a trained model give a wrong answer, often with changes a person would barely notice. Using such inputs at run time is called evasion.
- Learn it in: Data poisoning, backdoors and adversarial examples
- Related: Data poisoning, Backdoor attack
- Agent (also: AI agent)
- An AI system that decides its own next steps in a loop: it calls tools, reads the results and continues until the task is done or it must stop.
- The difference from a fixed workflow is who decides the path. In a workflow, your code decides; in an agent, the model does. That flexibility costs predictability, so agents need limits and logs.
- Learn it in: AI careers in India, Choosing your path, Automation with n8n: workflows and AI agents without heavy coding, Vibe coding with Claude Code: build real apps by describing them, How agents fail, Logs and traces for agents, The agent loop, Human approval steps, Tool calling, Multi-agent systems
- Related: Agent loop, Tool calling, Workflow
- Agent harness (also: Harness, Agent scaffold)
- Everything around the model that turns it into a working agent: the loop, tools, context handling, files it keeps on disk, checks on its work, permissions and limits.
- A common way to put it is agent = model + harness. Two coding tools can use the same model and still behave very differently, because one has better tools, tests that push back on mistakes and a way to remember progress between sessions.
- Learn it in: Advanced harness engineering, Harnesses for long-running agents
- Related: Harness engineering, Agent loop, Agent, Context engineering
- Agent loop (also: ReAct loop)
- The cycle an agent repeats: think about the next step, call a tool, read the result, decide again.
- Learn it in: The agent loop, LangGraph: agents as graphs, Planning and reflection, Advanced harness engineering
- Related: Agent, ReAct
- Agent memory
- What an agent keeps beyond one model call: the current conversation and task state (short-term) and facts or past experiences stored for later (long-term).
- Learn it in: Agent memory, Context engineering
- Related: Agent, Context engineering
- Agent SDK
- A library from a model provider or framework maker that supplies the agent loop, tool handling, tracing and guardrails so you write less glue code.
- Learn it in: Agent SDKs and sub-agents
- Related: Agent, Tool calling
- AI bill of materials (also: AI-BOM, ML-BOM, AIBOM)
- A machine-readable list of the models, datasets and software an AI system is built from, with versions, sources, hashes and licences. CycloneDX is one standard format for it.
- Learn it in: Securing the AI supply chain, Run an AI security programme
- Related: AI supply chain, Model card
- AI inventory
- A maintained list of every AI system and AI tool a company uses or builds, with its owner, data, vendor, models and risk level, so nothing is secured by accident or missed.
- Learn it in: Run an AI security programme
- Related: Shadow AI, Risk log
- AI security metrics
- Numbers that show whether an AI security programme works, such as the share of AI systems reviewed, attack success rate in tests and days to fix findings.
- Learn it in: Run an AI security programme
- Related: Red teaming, AI inventory, KPI
- AI supply chain (also: ML supply chain)
- Everything an AI system depends on that its team did not build: pre-trained models, datasets, packages, tools and hosting. Each can carry malicious code or poisoned data.
- Learn it in: Securing the AI supply chain
- Related: AI bill of materials, Pickle deserialisation, Model card, Open-weight model
- AI-generated phishing
- Scam messages written or tailored with AI tools, so they are fluent, personalised and produced in many languages at scale. Spelling mistakes stop being a warning sign; checks on the sender and the request matter more.
- Learn it in: AI-powered attacks and how defenders respond
- Related: Voice cloning, Deepfake, Large language model
- Air-gapped
- A system with no connection to the internet or other outside networks, used for the most sensitive data. Updates and models arrive by controlled transfer.
- Learn it in: Deployment models
- Related: Self-hosting, Deployment model
- Alert fatigue
- When analysts receive so many alerts, most of them harmless, that they start to skim, ignore or close them, and real attacks slip through unnoticed.
- Learn it in: Logs and SIEM, Threat detection with MITRE ATT&CK
- Related: SIEM, Detection rule, SOC (security operations centre)
- Anomaly detection
- Finding events that differ from a learned picture of normal, such as a login from a new country at 3 a.m. Useful for unknown attacks, but unusual is not the same as malicious, so it produces many false positives.
- Learn it in: Machine learning for threat detection
- Related: False positive, Unsupervised learning, Concept drift
- API (also: application programming interface)
- A defined way for one program to ask another for data or actions over the network, usually by sending JSON over HTTP.
- Learn it in: Reading and shaping JSON, Webhooks, Call a web API and read JSON, Call an LLM API from code, Enterprise connectors
- Related: REST API, Webhook
- Array (also: list, dynamic array)
- An ordered sequence of items stored side by side, so any item can be read by its position in constant time.
- A Python list is a dynamic array. Reading item 500 is instant, but inserting at the front means shifting every item along.
- Learn it in: Arrays, strings and hash maps
- Related: Hash map, Linked list
- Artificial intelligence (also: AI)
- Software that does tasks we used to think needed human judgement, such as understanding language, recognising images or making predictions.
- AI is an umbrella term. A spam filter, a face unlock and a chat assistant are all AI, though they work in very different ways. Most AI today learns from examples rather than following rules a person wrote.
- Learn it in: A short history of AI, What AI, machine learning and a language model are
- Related: Machine learning, Large language model
- async and await (also: asyncio)
- Python keywords for code that waits on slow things, such as network calls, without blocking other work. One process can then serve many requests at once. Python's asyncio library runs this code.
- Learn it in: Python for production, Serving an AI app
- Related: Latency
- Attack surface
- All the places an attacker could try to get in or cause harm: open ports, login pages, APIs, file uploads, staff email, third-party code and, for AI apps, every prompt and tool.
- Learn it in: How cyber attacks work
- Related: Vulnerability, Threat actor
- Attention
- A mechanism that lets a model weigh how much each part of the input matters for each part of the output. The core of transformers.
- Learn it in: Transformers
- Related: Transformer
- Audit log
- A tamper-resistant record of who did what, when and to which data, kept to investigate incidents and prove compliance.
- Learn it in: Incident response basics, Logs and SIEM, Audit logs, Versioning prompts and models, Accountability for agent actions, Incident response for AI, Security questionnaires, Documentation and handover
- Related: Logging, Tracing
- Authentication (also: AuthN)
- Proving who you are, for example with a password, a one-time code, a passkey or a fingerprint.
- Learn it in: Identity and access
- Related: Authorisation, Multi-factor authentication
- Automation
- Making software do a repeated task without a person, from a simple rule ("when a form arrives, add a row") to a workflow with AI steps.
- Learn it in: AI and jobs, Choosing your path, Mapping a process, What automation is, Make, Zapier and Power Automate, Your first n8n workflow, RPA basics, Triage for email and tickets
- Related: Workflow, Webhook
B
- Back-of-the-envelope estimate (also: capacity estimate, napkin maths)
- A quick rough calculation of load, storage or cost from a few assumptions, used to check whether a design is the right size before building it.
- If 1 crore users each send 10 messages a day, that is 100 crore messages a day, or about 12,000 a second on average. That number shapes every choice that follows.
- Learn it in: How to approach system design, Design walkthrough: a URL shortener
- Related: Horizontal scaling, Throughput
- Back-pressure (also: Verification gate)
- Automatic checks that push back on an agent's work, such as tests, linters, type checks or an evaluator agent, so it must fix a failure before it can call a step done.
- Without back-pressure an agent tends to report success on work that does not run. With it, a failing test result goes straight back into the loop, and the agent keeps working until the check passes or a limit is reached.
- Learn it in: Advanced harness engineering
- Related: Verifier, Unit test, Agent loop
- Backdoor attack (also: trojan model, model backdoor)
- Hidden behaviour planted in a model, usually through poisoned training data, that appears only when a trigger such as a phrase or image patch is present. Without the trigger the model behaves normally.
- Learn it in: Data poisoning, backdoors and adversarial examples
- Related: Data poisoning, Adversarial example
- Backpropagation
- The method that works out how much each weight in a neural network contributed to the error, so gradient descent can adjust it.
- Learn it in: Neural networks, PyTorch
- Related: Gradient descent, Neural network
- Base model
- A model that has only been pre-trained to predict the next token. It continues text rather than following instructions, so it is not yet a helpful assistant.
- Learn it in: How a chat model is made: pre-training, SFT and preference tuning
- Related: Pre-training, Supervised fine-tuning (SFT), Next-token prediction
- Baseline
- The measured value of a KPI before your system starts, so improvement can be proved.
- Learn it in: Measuring return on investment, Discovery, Talking to executives, The KPI that matters
- Related: KPI
- Batch API (also: batch processing)
- A way to send many model requests at once for processing within hours instead of seconds, usually at a large discount.
- Learn it in: Batching and model tiering, Cost control
- Related: Cost per request, Model tiering
- Batch inference
- Running a model on many records at once on a schedule, such as nightly scores for all customers, rather than one request at a time.
- Learn it in: Serving a model
- Related: Inference
- Batch size
- How many examples a model processes before each weight update during training.
- Learn it in: PyTorch, Fine-tuning in practice: tools, settings and reading the loss curve
- Related: Epoch, Gradient descent
- Benchmark
- A fixed public test set used to compare models on one skill, such as maths, coding or general knowledge. Useful for a shortlist, but it does not measure your task.
- Learn it in: Choosing a model: benchmarks, leaderboards and your own tests
- Related: Evals, Data contamination, Leaderboard
- Best-of-N (also: best of N, self-consistency)
- Generating several answers to the same question and keeping the best, chosen by a checker, a scoring model or a majority vote. It costs roughly N times as much.
- Learn it in: Reasoning models and test-time compute: when to let the model think longer
- Related: Test-time compute, Verifier
- Bias
- When an AI system's outputs are systematically worse or unfair for some groups, often because of skewed training data.
- Learn it in: Bias and fairness, Statistics for machine learning, LLM judges and their limits
- Related: Responsible AI
- Big-O notation (also: Big O, time complexity, space complexity)
- A way to describe how an algorithm's running time or memory grows as its input grows, ignoring constants. O(n) grows in step with the input; O(n²) grows with its square.
- If checking one list of 1,000 phone numbers takes a moment, an O(n) method on 10 lakh numbers takes about a thousand moments, but an O(n²) method takes about a million times as long. Big-O lets you see that before you run anything.
- Learn it in: Arrays, strings and hash maps, Big-O and complexity, Sorting, binary search and two pointers
- Related: Hash map, Binary search
- Binary search (also: bisection)
- A way to find an item in a sorted list by checking the middle and discarding the half that cannot hold it, so each step halves what is left.
- Finding a word in a printed dictionary works this way. Binary search finds one roll number among 10 lakh sorted ones in about 20 checks.
- Learn it in: Big-O and complexity, Sorting, binary search and two pointers
- Related: Big-O notation, Two pointers
- Binary tree (also: binary search tree, BST, tree)
- A tree where each node has at most two children. In a binary search tree, smaller values go left and larger go right, so a lookup halves the search each step.
- Folders on your laptop form a tree. A balanced binary search tree of 10 lakh items finds any item in about 20 steps.
- Learn it in: Trees and graphs
- Related: Graph, Depth-first search
- Branch
- In Git, a separate line of work. You change code on a branch, then merge it back when it is reviewed, so the main line stays working.
- Learn it in: Git branches, merging and rebasing, Git with a team
- Related: Pull request, Merge conflict
- Breadth-first search (also: BFS)
- A way to explore a graph level by level using a queue, visiting all neighbours of a node before going further. It finds the shortest path when every edge counts the same.
- To find the fewest metro stops from Rajiv Chowk, BFS checks every station one stop away, then two stops away, and so on.
- Learn it in: Trees and graphs
- Related: Depth-first search, Queue, Graph
- Breakpoint
- A marked line where a debugger pauses the program, so you can look at its state at that moment.
- Learn it in: Debugging systematically
- Related: Debugger
- Budget limit (also: spend cap)
- A cap on how many steps, tokens or rupees an agent or feature may use per task, user or day.
- Learn it in: Errors, retries and cost, How agents fail, Loop limits, budgets and kill switches, LLM gateways, Cost control, Harnesses for long-running agents
- Related: Kill switch, Cost per request
C
- Cache invalidation (also: cache expiry, TTL)
- Deciding when a cached copy is out of date and must be removed or refreshed, for example after a set time-to-live or when the original data changes.
- If a price changes but the cached page still shows the old one, shoppers see the wrong price. A short TTL or clearing the cache on update fixes it.
- Learn it in: Caching and CDNs
- Related: Caching, Content delivery network
- Caching (also: prompt caching)
- Keeping a result so the same request can be answered again without redoing the work. Prompt caching reuses a long, unchanged prompt prefix.
- Learn it in: Pricing AI features, Caching and CDNs, Long context or RAG, Design walkthrough: a URL shortener, Caching and speed
- Related: Latency, Cost per request
- Catastrophic forgetting
- When fine-tuning a model on one task makes it worse at things it could do before. It is why a fine-tuned model needs tests on old tasks as well as the new one.
- Learn it in: How fine-tuning works: data, training and what can go wrong, Model merging: combining fine-tuned models into one
- Related: Fine-tuning, Overfitting, Training
- Change data capture (also: CDC)
- Streaming every insert, update and delete from a database as it happens, so other systems such as a search index stay in sync.
- Learn it in: Enterprise connectors, Data pipelines for AI
- Related: ETL and ELT
- Change management
- Helping people adopt a new way of working: explaining why, training, supporting and reinforcing, and dealing with resistance.
- Learn it in: Change management and training
- Related: Adoption
- Change request
- A formal proposal to change agreed scope, stating the impact on time, cost and other work, for the client to approve or decline.
- Learn it in: Managing scope creep
- Related: Scope creep
- Character encoding (also: UTF-8)
- The rule a file uses to store letters as bytes. UTF-8 handles every script, including Tamil and Devanagari; older files may use other encodings and show garbled text if read wrongly.
- Learn it in: Files, tables and PDFs
- Related: Document parsing
- Checkpoint
- A saved snapshot of an agent's state, so it can resume after a pause, a crash or a person's approval, or be replayed for debugging.
- Learn it in: Agent memory, Human approval steps, LangGraph: agents as graphs, Harnesses for long-running agents
- Related: LangGraph, Human in the loop
- Chunk overlap
- Repeating a little text at the end of one chunk and the start of the next, so a sentence split at the boundary is not lost.
- Learn it in: Chunk documents for retrieval
- Related: Chunking
- Chunking
- Splitting documents into smaller passages before embedding them, so search can return the exact part that answers a question.
- Learn it in: Chunk documents for retrieval, Parsing documents and OCR, Metadata that helps retrieval, Data pipelines for AI
- Related: Chunk overlap, Embedding, Retrieval-augmented generation
- CI security tests
- Security checks that run automatically on every code change in a CI pipeline, such as a fixed set of attack prompts and dependency scans, and fail the build when a threshold is crossed.
- Learn it in: Security tests in CI for AI apps
- Related: CI/CD, Regression test, Red teaming
- CI/CD
- Continuous integration and continuous delivery: automatically testing every change and deploying it through stages when tests pass.
- Learn it in: CI/CD with GitHub Actions, Versioning prompts and models, Regression tests for prompts, Security tests in CI for AI apps
- Related: Regression test, Environment
- CIA triad (also: Confidentiality, integrity and availability)
- The three goals of security: confidentiality (only the right people see data), integrity (data is not changed without permission) and availability (systems work when needed).
- Learn it in: How cyber attacks work
- Related: Threat actor, Attack surface
- Citation
- A pointer from a statement in an answer to the passage that supports it, so a reader can check it.
- Learn it in: AI for research and analysis, Checking an AI answer, UX for uncertain output, Grounding, citations and refusing to guess, Build a RAG app, Metadata that helps retrieval
- Related: Grounding, Retrieval-augmented generation
- Classification
- Predicting a category, such as fraud or not fraud, or which of five product types.
- Learn it in: Classification
- Related: Regression, Precision and recall
- Cloud misconfiguration
- A cloud setting left unsafe by the customer, such as a public storage bucket, an over-broad access policy, an open database port or a leaked access key. A common cause of cloud data exposure.
- Learn it in: Cloud security basics
- Related: Shared responsibility model, Identity and access management, Secrets management
- Clustering (also: k-means)
- Grouping examples so that those in the same group are similar, without predefined labels, such as customer segments. k-means is a common method.
- Learn it in: Unsupervised learning
- Related: Unsupervised learning
- Code review
- A teammate reading your changes before they merge, to catch bugs, unclear code and risky decisions.
- Learn it in: Git with a team, Code review: giving and getting good reviews
- Related: Pull request
- Code smell
- A sign that code may be hard to change: very long functions, duplicated logic, unclear names, or one change needing edits in many places.
- Learn it in: Clean code and refactoring
- Related: Refactoring
- Commit
- A saved snapshot of changes in Git, with a message explaining what changed and why.
- Learn it in: Undoing things in Git: restore, reset, revert and reflog, The terminal and Git
- Related: Git
- Compute-optimal (also: Chinchilla-optimal)
- The split between model size and training tokens that gives the best quality for a fixed amount of training compute. The Chinchilla rule of thumb is about 20 training tokens per parameter.
- Learn it in: Model size and scaling laws: parameters, training data and compute
- Related: Scaling law, Parameters
- Computer-use agent (also: browser agent, GUI agent)
- An agent that operates software the way a person does, by looking at screenshots and clicking and typing, for apps that have no API.
- Learn it in: RPA basics, Browser and computer-use agents
- Related: Agent, Sandbox
- Concept drift
- When the relationship between inputs and the right answer changes over time, such as fraud patterns shifting, so a model's accuracy falls.
- Learn it in: Machine learning for threat detection, Monitoring and drift
- Related: Drift, Monitoring
- Confused deputy (also: confused deputy problem)
- When a program with more authority than its caller is tricked into using that authority on the caller's behalf, such as an agent with admin access reading files for a user who cannot open them.
- Learn it in: Securing agent tools and MCP servers
- Related: Least privilege, Excessive agency, Tool poisoning
- Confusion matrix
- A table counting a classifier's true positives, false positives, true negatives and false negatives.
- Learn it in: Choosing a metric, Classification
- Related: Precision and recall
- Connector
- Code that pulls data from, or pushes data to, a business system such as SharePoint, Salesforce or SAP, handling its API, sign-in and paging.
- Learn it in: Enterprise connectors
- Related: ETL and ELT, API
- Container
- A packaged, isolated process that carries its own code, libraries and settings, so it runs the same on a laptop, a server or the cloud.
- Learn it in: One cloud in depth, Docker and Compose, Kubernetes and Helm basics
- Related: Image, Docker Compose
- Content delivery network (also: CDN, edge cache)
- A network of servers in many cities that keeps copies of files such as images and videos close to users, so they load faster and the main server does less work.
- A learner in Chennai gets a course video from a CDN server in Chennai instead of from a server in the US, so it starts quickly.
- Learn it in: Caching and CDNs
- Related: Caching, Cache invalidation
- Context compaction (also: Compaction)
- Summarising an agent's history in place when its context window nears the limit, so the same session can continue with a shorter record of the goal, decisions and open items.
- Compaction keeps continuity but never gives a clean slate. The other option is a context reset: end the session, write a structured handoff to a file, and start a fresh agent that reads it.
- Learn it in: Harnesses for long-running agents
- Related: Context window, Context engineering, Agent memory
- Context engineering
- Choosing what goes into the model's context window on each call, including instructions, retrieved data, memory and tool results, and what stays out.
- Learn it in: Agent SDKs and sub-agents, Context engineering, Advanced harness engineering
- Related: Context window, Retrieval-augmented generation, Agent
- Context window
- The most tokens a model can take in one call, counting your instructions, any documents, the conversation so far and its own reply.
- Learn it in: Prompting without jargon, Use a chat assistant for a real task and check the answer, What AI cannot do, Tokens and the context window, Long context or RAG, Agent memory, Long documents without vectors, Context engineering
- Related: Token, Context engineering, Retrieval-augmented generation
- Continuous batching (also: in-flight batching)
- Serving many requests together on one GPU and swapping a finished request out for a new one straight away, instead of waiting for the whole batch. It raises throughput a lot.
- Learn it in: Inside LLM inference: prefill, decode and the KV cache
- Related: Throughput, Batch size
- Convolutional neural network (also: CNN)
- A neural network that scans images with small learned filters to detect edges, shapes and objects. The classic architecture for vision.
- Learn it in: CNNs for images
- Related: Transfer learning, Deep learning
- Cosine similarity
- A score from -1 to 1 for how closely two embeddings point in the same direction. Higher means more similar in meaning.
- Learn it in: Embeddings, Vector search
- Related: Embedding, Semantic search
- Cost per request
- What one user request costs in model tokens, search and infrastructure. The number that decides whether an AI feature can pay for itself.
- Learn it in: AI PM interviews, Pricing AI features, Tokens and the context window, Tracing LLM apps, Loop limits, budgets and kill switches, Cost of ownership, Batching and model tiering, Cost control, The design interview, Quality, cost and speed trade-offs
- Related: Token, Caching, Model tiering
- Cross-encoder
- A model that reads a query and a passage together and scores how well the passage answers it. Accurate but slow, so it is used to rerank a short list.
- Learn it in: Hybrid search, reranking and filters
- Related: Reranking, Embedding
- Cross-entropy (also: cross-entropy loss, log loss)
- The loss a language model is trained to lower: how far its predicted probabilities are from the tokens that actually came next. Perplexity is its exponential.
- Learn it in: Perplexity and cross-entropy: how language models are measured
- Related: Perplexity, Loss function
- Cross-site scripting (also: XSS)
- A web flaw where an attacker gets their script to run in other users' browsers, usually because the site shows user input without encoding it for HTML.
- Learn it in: Web app security and the OWASP Top 10
- Related: OWASP Top 10, SQL injection
- Cross-validation (also: k-fold)
- Testing a model several times on different held-out parts of the data and averaging, for a more reliable estimate of performance.
- Learn it in: Cross-validation and leakage, Feature engineering
- Related: Overfitting, Data leakage
- CSV
- Comma-separated values: a plain text file of rows and columns, the most common way to move tables between programs.
- Learn it in: Files and errors, Tables with pandas
- Related: DataFrame
- CVE (also: Common Vulnerabilities and Exposures)
- A public ID for a known vulnerability, such as CVE-2021-44228, so everyone can refer to the same flaw. Severity scores and fix details are added by databases such as NVD.
- Learn it in: How cyber attacks work
- Related: Vulnerability, Exploit
D
- Data contamination (also: benchmark contamination)
- When a model was trained on the questions of a test it is later scored on, so the score shows memory rather than skill. Also called benchmark leakage.
- Learn it in: Choosing a model: benchmarks, leaderboards and your own tests
- Related: Benchmark, Evals, Overfitting
- Data curation
- Choosing, cleaning, labelling and balancing the examples a model is trained on, judged on quality, coverage and quantity.
- Learn it in: Building a training dataset: quality, coverage and quantity
- Related: Training, Fine-tuning, Synthetic data
- Data exfiltration
- Getting private data out of a system without permission, for example by tricking an AI assistant into putting it in a link or reply.
- Learn it in: Threats to LLM apps: the OWASP Top 10, Permissions and sandboxing, Threat model an AI system, Defend against prompt injection, Red-team an AI system, Run an AI security programme
- Related: Prompt injection, Personal data
- Data fiduciary
- Under India's DPDP Act, the organisation that decides why and how personal data is processed, and carries the legal duties for it.
- Learn it in: Compliance for AI systems
- Related: DPDP Act, Data processor
- Data leakage (also: leakage)
- When information that would not be available at prediction time sneaks into training, making a model look much better in testing than it will be in real use.
- Learn it in: Cross-validation and leakage, Feature engineering, ML interviews
- Related: Cross-validation
- Data masking (also: redaction, pseudonymisation)
- Replacing sensitive values with placeholders or fake values, such as Aadhaar numbers with XXXX-XXXX-1234, so data can be used without exposing people.
- Learn it in: Data quality and personal data
- Related: Personal data, DPDP Act
- Data parallelism (also: replica parallelism)
- Running a full copy of the model on each GPU and splitting the work between them: different training batches, or different user requests when serving (replicas).
- Learn it in: Running models on many GPUs: replicas, tensor and pipeline parallelism
- Related: Tensor parallelism, Pipeline parallelism
- Data poisoning (also: training data poisoning, model poisoning)
- Planting or changing data that a model learns from, such as training, fine-tuning or knowledge-base data, so the model makes the mistakes an attacker wants.
- Learn it in: Data poisoning, backdoors and adversarial examples
- Related: Backdoor attack, Adversarial example, Fine-tuning, Retrieval-augmented generation
- Data processor
- An organisation that processes personal data on behalf of a data fiduciary (or a controller under GDPR), such as a vendor running an AI system for a bank.
- Learn it in: Compliance for AI systems
- Related: Data fiduciary, DPDP Act
- Data residency
- A requirement that data is stored and processed in a particular country or region, often set by law, regulators or the client's contracts.
- Learn it in: One cloud in depth, Deployment models
- Related: Region, DPDP Act
- Data versioning (also: DVC)
- Keeping versions of datasets like code, so any model can be traced to and retrained on the exact data it used. DVC is a common tool for this.
- Learn it in: Data versioning, Experiment tracking and model registry
- Related: Experiment tracking
- DataFrame (also: pandas)
- A table of rows and named columns in pandas, the Python library for working with tabular data.
- Learn it in: Tables with pandas
- Related: CSV
- Debugger
- A tool that pauses a running program so you can inspect variables, step line by line and see exactly where it goes wrong. In Python: breakpoint() and pdb, or your editor's debugger.
- Learn it in: Debugging systematically
- Related: Breakpoint
- Decision tree
- A model that makes predictions by asking a series of yes-or-no questions about features, like a flowchart learned from data.
- Learn it in: Trees and gradient boosting
- Related: Random forest, Gradient boosting
- Deep learning
- Machine learning with neural networks of many layers, which powers modern image, speech and language models.
- Learn it in: Neural networks
- Related: Neural network
- Deepfake
- An AI-made or AI-altered video, image or voice recording that makes a real person appear to say or do something they did not.
- Learn it in: Deepfakes and misinformation, Images, voice and other AI tools, AI-powered attacks and how defenders respond
- Related: Generative AI, Misinformation
- Dependency injection
- Passing a part a function needs, such as a database or model client, in from outside instead of creating it inside, so it can be swapped for a fake in tests.
- Learn it in: Design principles: modules, SOLID and dependency injection
- Related: SOLID principles, Mocking
- Deployment model
- Where an AI system runs and who controls it: a vendor's cloud API, the client's own cloud account, their data centre, or an air-gapped site.
- Learn it in: Feasibility and build versus buy, Deployment models
- Related: Self-hosting, Air-gapped
- Depth-first search (also: DFS)
- A way to explore a graph by going as deep as possible along one branch before backing up, using recursion or a stack.
- DFS is how you would search a maze by following one corridor to its end before trying the next. It is used to find connected groups and detect cycles.
- Learn it in: Trees and graphs
- Related: Breadth-first search, Stack, Recursion
- Detection rule
- A saved query or condition over logs that raises an alert when it matches suspicious activity, such as twenty failed logins followed by a success from the same address.
- Learn it in: Logs and SIEM, Threat detection with MITRE ATT&CK
- Related: Sigma rule, SIEM, MITRE ATT&CK
- Dictionary (also: dict)
- A Python collection of key-value pairs, such as a phone book from names to numbers, for looking things up by key.
- Learn it in: Variables, loops and functions in Python, Lists and dictionaries
- Related: List, JSON
- Diffusion model
- A generative model that creates images or video by starting from noise and removing it step by step, guided by a prompt.
- Learn it in: AI image generation and editing with Gemini (Nano Banana) and Midjourney, Create cinematic AI videos with Google Veo and Flow
- Related: Text-to-video
- Dimensionality reduction (also: PCA)
- Compressing many features into a few that keep most of the information, for visualisation or simpler models. PCA is a common method.
- Learn it in: Unsupervised learning
- Related: Unsupervised learning
- Discovery
- The first phase of a client engagement: understanding the problem, the people, the process and the constraints before proposing a solution.
- Learn it in: Mapping a process, Choosing a problem for AI, Discovery, A case study, end to end, The FDE interview loop
- Related: KPI, Scope of work
- Distillation (also: knowledge distillation)
- Training a small student model to copy the answers of a large teacher model, to get similar quality on a task at lower cost and higher speed. Also called knowledge distillation.
- Learn it in: Synthetic data and distillation: training with model-made data
- Related: Synthetic data, Fine-tuning
- DNS (also: Domain Name System)
- The internet's phonebook: it turns names like example.com into IP addresses. Defenders watch DNS because malware often looks up its control servers by name.
- Learn it in: Networking for security
- Related: TCP/IP, Firewall
- Docker Compose (also: Compose)
- A tool that starts several containers together from one file, such as your API, a database and a vector store, with their settings and links.
- Learn it in: Docker and Compose
- Related: Container
- Dockerfile
- A text file of steps that builds a container image: start from a base, copy files, install packages, set the start command.
- Learn it in: Docker and Compose
- Related: Image
- Document parsing
- Pulling clean text, tables and structure out of files such as PDFs, Word documents and slides so software can use them.
- Learn it in: Files, tables and PDFs, Extracting data from documents, Parsing documents and OCR, Multimodal RAG
- Related: OCR, Chunking
- DPDP Act (also: Digital Personal Data Protection Act)
- India's Digital Personal Data Protection Act, 2023: the law on how organisations collect, use and protect people's personal data.
- Learn it in: AI rules in India and abroad, Your company's AI policy, Use AI safely at work, Data strategy for AI products, Writing an AI policy, Data quality and personal data, Compliance for AI systems, Model extraction and privacy attacks, Run an AI security programme
- Related: Personal data, GDPR
- DPO (direct preference optimisation)
- A simpler alternative to RLHF that learns from the same pairs of preferred and rejected answers, but trains the model on them directly, with no separate reward model.
- Learn it in: How a chat model is made: pre-training, SFT and preference tuning
- Related: RLHF (reinforcement learning from human feedback), Reward model
- Drift
- When live data or behaviour slowly changes from what a system was built and tested on, so its quality drops without any code change.
- Learn it in: Monitoring and drift, Online evals and feedback
- Related: Monitoring, Evals
- Durable execution
- Running a long job so that each completed step is saved, and after a crash or restart the job resumes from the last saved step instead of starting again.
- For agents this means saving the state after each tool call or task, with step ids so a step is never done twice. A run that takes hours then survives a laptop going to sleep or a server restart.
- Learn it in: Harnesses for long-running agents
- Related: Checkpoint, Idempotency, Agent loop
- Dynamic programming (also: DP)
- Solving a problem by breaking it into overlapping smaller problems, solving each once and reusing the answers instead of recomputing them.
- Working out the fewest coins for ₹93 is easier if you already know the answers for ₹92, ₹88 and ₹83. DP builds those answers up once and looks them up.
- Learn it in: Recursion and dynamic programming
- Related: Memoization, Recursion
E
- Embedding (also: vector embedding)
- A list of numbers that represents the meaning of a piece of text, so texts with similar meanings get similar numbers.
- If you embed "car" and "automobile", the two lists of numbers point in nearly the same direction, even though the words share no letters. That is what lets search find meaning rather than exact words.
- Learn it in: Chunk documents for retrieval, Embeddings, Transformers, Data pipelines for AI
- Related: Vector database, Cosine similarity, Semantic search
- Encryption
- Scrambling data with a key so only someone with the right key can read it. Symmetric encryption uses one shared key; asymmetric uses a public and private key pair.
- Learn it in: Cryptography basics
- Related: Hashing, TLS
- Endpoint
- One URL and method of an API that does one thing, such as POST /tickets to create a ticket.
- Learn it in: REST APIs with FastAPI
- Related: REST API, API
- Environment (also: staging, production)
- A separate copy of a system for one purpose, such as development, staging for testing and production for real users, each with its own settings and data.
- Learn it in: CI/CD with GitHub Actions, Environments, secrets and Terraform
- Related: CI/CD, Environment variable
- Environment variable (also: env var)
- A named setting a program reads from its surroundings, such as DATABASE_URL or an API key, so the same code runs in different places without edits.
- Learn it in: Cryptography basics, Linux, the shell and SSH, Call an LLM API from code, Environments, secrets and Terraform
- Related: Secrets management
- Epoch
- One complete pass through the training data during training.
- Learn it in: PyTorch, Fine-tuning in practice: tools, settings and reading the loss curve
- Related: Batch size
- Escalation
- Raising a problem to someone with more authority when it cannot be solved at the current level in time.
- Learn it in: Demos, risk logs and escalation
- Related: Risk log
- ETL and ELT
- Ways to move data between systems. ETL transforms data before loading it into the destination; ELT loads raw data first and transforms it there.
- Learn it in: Data versioning, Data pipelines for AI
- Related: Change data capture, Connector
- EU AI Act
- The European Union's AI law. It sorts AI uses by risk, bans a few, and puts duties such as testing, records and human oversight on high-risk ones.
- Learn it in: AI rules in India and abroad, AI risk frameworks
- Related: Responsible AI, GDPR
- Evals (also: evaluations)
- Repeatable tests that score an AI system's outputs on a fixed set of cases, so you know whether a change made it better or worse.
- Learn it in: AI careers in India, Bias and fairness, Feasibility and build versus buy, Choosing a problem for AI, How AI products work, Statistics for machine learning, AI PM interviews, System prompts and prompt design, Choosing a model: benchmarks, leaderboards and your own tests, Test sets and evals, Evaluating agents, LLM judges and their limits, Online evals and feedback, Regression tests for prompts, Scoring open-ended answers: exact checks, similarity and side-by-side comparison, Design an AI system, The design interview, Documentation and handover, The KPI that matters, Portfolio and job search, Prompting, RAG or fine-tuning, Quality, cost and speed trade-offs
- Related: Golden set, LLM as judge, Regression test
- Exact match
- Scoring an answer right only if it equals the reference answer, often after small clean-ups such as lowercasing. It suits short, closed answers like numbers and labels.
- Learn it in: Scoring open-ended answers: exact checks, similarity and side-by-side comparison
- Related: Evals, Semantic similarity
- Exception (also: error handling, try/except)
- An error that happens while a program runs, such as a missing file or text that cannot become a number. Code can catch and handle it.
- Learn it in: Files and errors
- Related: Function
- Excessive agency
- When an LLM app or agent has more tools, permissions or freedom to act than its task needs, so a mistake or a prompt injection can cause real damage. Fix it with least privilege and human approval.
- Learn it in: Threats to LLM apps: the OWASP Top 10, Securing agent tools and MCP servers
- Related: Least privilege, Human in the loop, OWASP Top 10 for LLM Applications
- Experiment tracking
- Recording every training run's code version, data, parameters, metrics and model, so results can be compared and reproduced.
- Learn it in: Data versioning, Experiment tracking and model registry
- Related: Model registry
- Exploit
- Code or a technique that takes advantage of a vulnerability to make a system do something it should not.
- Learn it in: How cyber attacks work
- Related: Vulnerability, CVE
F
- Factual consistency
- Whether every claim in an answer is supported by the source it should rely on, such as a document, a policy or search results.
- Learn it in: Catching hallucinations: why models make things up and how to detect it
- Related: Hallucination, Faithfulness, Grounding
- Faithfulness (also: groundedness)
- How far an answer's claims are supported by the retrieved sources. A faithful answer adds nothing the sources do not say.
- Learn it in: Check whether retrieval works, Grounding, citations and refusing to guess, Catching hallucinations: why models make things up and how to detect it
- Related: Grounding, Hallucination, Evals
- Fallback model
- A second model a system switches to when the first is down, rate-limited or too slow.
- Learn it in: LLM gateways
- Related: LLM gateway
- False positive
- An alert or prediction that says something is bad when it is not, such as a genuine login flagged as an attack. Too many false positives waste analyst time and teach people to ignore alerts.
- Learn it in: Machine learning for threat detection
- Related: Precision and recall, Confusion matrix, Anomaly detection
- Feature
- An input variable a model uses, such as age, monthly income or number of past orders.
- Learn it in: NumPy and plotting, Feature engineering, Regression
- Related: Label, Feature engineering
- Feature engineering
- Creating useful inputs for a model from raw data, such as days since last order or average spend per month.
- Learn it in: Cross-validation and leakage, Feature engineering, Unsupervised learning
- Related: Feature
- Few-shot prompting (also: in-context examples)
- Putting a few worked examples of the task in the prompt so the model copies the pattern.
- Learn it in: Prompting without jargon, System prompts and prompt design
- Related: Prompt, Prompt engineering
- Fine-tuning
- Training an existing model further on your own examples so it learns a style, format or narrow task.
- Learn it in: CNNs for images, Building a training dataset: quality, coverage and quantity, Data poisoning, backdoors and adversarial examples, How fine-tuning works: data, training and what can go wrong, Synthetic data and distillation: training with model-made data, Model merging: combining fine-tuned models into one, Prompting, RAG or fine-tuning
- Related: Training, Retrieval-augmented generation, Prompt engineering
- Firewall
- A system that allows or blocks network traffic by rules, such as source, destination, port and protocol. A web application firewall (WAF) inspects web requests instead.
- Learn it in: Networking for security
- Related: Port, TCP/IP
- Fixture
- In pytest, a reusable piece of test setup, such as a temporary database or a sample file, that tests ask for by name.
- Learn it in: Project tooling: environments, tests and logs
- Related: pytest
- Function
- A named, reusable block of code that takes inputs, does a job and returns a result.
- Learn it in: Variables, loops and functions in Python, Files and errors, Write your own functions
- Related: Variable
G
- garak
- An open-source LLM vulnerability scanner from NVIDIA that sends many ready-made attack probes, such as jailbreaks and injections, to a model and reports which ones got through.
- Learn it in: Red-team an AI system
- Related: Red teaming, PyRIT, Jailbreak
- GDPR (also: General Data Protection Regulation)
- The European Union's data protection law. It also covers companies outside the EU, including Indian ones, that offer goods or services to people in the EU or monitor their behaviour.
- Learn it in: Compliance for AI systems
- Related: DPDP Act, Personal data
- Generative AI (also: GenAI)
- AI that creates new content, such as text, images, audio, video or code, from a description or example, rather than only sorting or predicting.
- Learn it in: AI and jobs, AI for writing and summarising, Deepfakes and misinformation, A short history of AI, Images, voice and other AI tools
- Related: Large language model, Artificial intelligence
- Git
- A version control tool that records the history of changes to files, so you can review, undo and share work.
- Learn it in: The terminal and Git, Harnesses for long-running agents
- Related: Branch, Pull request, Terminal
- Git reset
- Moves your branch back to an earlier commit. Soft keeps your changes staged, mixed keeps them unstaged, hard throws them away. Use it only on commits nobody else has.
- Learn it in: Undoing things in Git: restore, reset, revert and reflog
- Related: Git revert, Commit
- Git revert
- A new commit that undoes an earlier one. Safe on shared branches, because it adds history instead of rewriting it.
- Learn it in: Undoing things in Git: restore, reset, revert and reflog
- Related: Commit, Git reset
- Golden set (also: test set, eval set)
- A fixed list of test inputs with known good answers or sources, used to score a system the same way every time.
- Learn it in: Data strategy for AI products, Check whether retrieval works, Test sets and evals, Evaluating agents, LLM judges and their limits, Regression tests for prompts
- Related: Evals, Recall at k
- GPU
- A graphics processing unit: the chip that runs AI models fast because it does many calculations at once. Its memory limits which models fit.
- Learn it in: GPUs and quantization, LoRA, QLoRA and the GPU memory maths, Running models on many GPUs: replicas, tensor and pipeline parallelism
- Related: vLLM, Quantization
- Gradient accumulation
- Adding up the learning signal from several small batches before updating the weights, so training behaves as if it used a bigger batch without needing more memory.
- Learn it in: Fine-tuning in practice: tools, settings and reading the loss curve
- Related: Batch size, Learning rate
- Gradient boosting (also: XGBoost, LightGBM)
- Building trees one after another, each correcting the errors of those before. XGBoost and LightGBM are popular versions, strong on tabular data.
- Learn it in: Trees and gradient boosting
- Related: Decision tree, Random forest
- Gradient descent
- The method most models use to learn: repeatedly nudge the weights in the direction that reduces the error a little, until it stops improving.
- Learn it in: Maths for machine learning, Neural networks
- Related: Loss function, Backpropagation
- Graph (also: graph data structure, nodes and edges, adjacency list)
- A set of nodes joined by edges, used to model networks such as roads, friendships or links between web pages.
- Delhi Metro is a graph: stations are nodes and the tracks between them are edges. Finding the fewest stops between two stations is a graph search.
- Learn it in: Trees and graphs
- Related: Breadth-first search, Depth-first search, Binary tree
- Graph RAG (also: GraphRAG)
- Retrieval that uses a knowledge graph built from documents, so answers can draw on relationships and on summaries of whole topics, not only similar passages.
- Learn it in: Graph RAG
- Related: Knowledge graph, Retrieval-augmented generation
- Grounding
- Keeping a model's answer tied to supplied sources, and having it say so when the sources do not contain the answer.
- Learn it in: AI for research and analysis, UX for uncertain output, Chat assistants for websites and WhatsApp, Grounding, citations and refusing to guess, Build a RAG app, Catching hallucinations: why models make things up and how to detect it
- Related: Retrieval-augmented generation, Citation, Hallucination
- Guardrails
- Checks around a model that block or fix unsafe, off-policy or malformed inputs and outputs.
- Learn it in: How AI products work, Triage for email and tickets, Data quality and personal data, How agents fail, Guardrails, Defend against prompt injection, Grow an AI app in layers: context, guardrails, routing, caching and agents, Design an AI system
- Related: Prompt injection, Human in the loop
H
- Hallucination (also: confabulation)
- When a model states something false or invents a source with full confidence.
- It happens because the model writes what sounds likely, not what it has checked. Grounding answers in retrieved documents and asking for citations reduces it; checking important claims is still your job.
- Learn it in: AI for research and analysis, AI for writing and summarising, Checking an AI answer, Use AI safely at work, Use a chat assistant for a real task and check the answer, What AI cannot do, What AI, machine learning and a language model are, UX for uncertain output, Grounding, citations and refusing to guess, LLMs in the SOC, Catching hallucinations: why models make things up and how to detect it
- Related: Grounding, Citation, Large language model
- Handover
- Transferring a system to the people who will own it, with documentation, training and a period of support.
- Learn it in: Documentation and handover, A case study, end to end, Portfolio and job search
- Related: Runbook
- Harness engineering
- Improving an agent by changing what surrounds the model, such as tools, context handling, progress files, tests and stop rules, and measuring each change on a fixed set of tasks.
- It follows prompt engineering and context engineering. When an agent makes a mistake, you add something to the harness, such as a rule file, a test or a better tool, so it does not make that mistake again. Then you check on the same tasks that the change really helped.
- Learn it in: Advanced harness engineering
- Related: Agent harness, Context engineering, Prompt engineering, Evals
- Hash map (also: hash table, dictionary, dict)
- A data structure that stores key and value pairs and finds a value from its key in about the same time no matter how many pairs it holds.
- Python's dict is a hash map. Looking up a customer by phone number in a dict of 10 lakh customers is as quick as in a dict of ten, because the key is turned into a position with a hash function.
- Learn it in: Arrays, strings and hash maps, Big-O and complexity
- Related: Big-O notation, Array
- Hashing (also: Hash function)
- Turning data into a fixed-length fingerprint that cannot be reversed. Used to check integrity and, with slow algorithms such as Argon2 or bcrypt, to store passwords.
- Learn it in: Cryptography basics
- Related: Encryption, TLS
- Helm
- A package manager for Kubernetes. A Helm chart bundles all the settings to install an app, with values you change per environment.
- Learn it in: Kubernetes and Helm basics
- Related: Kubernetes
- HIPAA
- The US law protecting health information. Indian vendors that handle US patients' data for US healthcare clients usually must follow its rules under a business associate agreement.
- Learn it in: Compliance for AI systems
- Related: GDPR, DPDP Act
- Horizontal scaling (also: scaling out, vertical scaling, scaling up)
- Handling more load by adding more machines rather than making one machine bigger. Making one machine bigger is vertical scaling.
- When a sale brings ten times the usual shoppers, a site that scales horizontally adds more app servers behind a load balancer instead of buying one giant server.
- Learn it in: Scaling and load balancing
- Related: Load balancer, Throughput
- HTTP method
- The verb of a web request: GET reads, POST creates, PUT or PATCH changes, DELETE removes.
- Learn it in: REST APIs with FastAPI
- Related: REST API, HTTP status code
- HTTP status code
- The three-digit number a server sends back: 2xx means it worked, 4xx means the request was wrong, 5xx means the server failed.
- Learn it in: Call a web API and read JSON, Errors, retries and cost, REST APIs with FastAPI
- Related: REST API, Rate limit
- Human in the loop (also: HITL)
- A design where a person reviews or approves an AI system's output before it takes effect, at least for risky cases.
- Learn it in: AI and jobs, What AI cannot do, Choosing a problem for AI, Designing human in the loop, UX for uncertain output, AI PM interviews, Chat assistants for websites and WhatsApp, Extracting data from documents, Triage for email and tickets, LLMs in the SOC, How agents fail, Human approval steps, LangGraph: agents as graphs, Accountability for agent actions, Securing agent tools and MCP servers, Browser and computer-use agents, Harnesses for long-running agents, Design an AI system, Change management and training, Discovery
- Related: Guardrails, Agent
- Hybrid search
- Running keyword search and semantic search together and merging the results, so you catch both exact terms and paraphrases.
- Learn it in: Hybrid search, reranking and filters
- Related: Keyword search, Semantic search, Reranking
- HyDE
- Hypothetical document embeddings: asking a model to write a likely answer, then searching with that answer's embedding, because answers look more like documents than questions do.
- Learn it in: Query rewriting, HyDE and parent-child chunks
- Related: Query rewriting, Embedding
I
- Idempotency
- A property of an action that gives the same result whether it runs once or several times, so retries cannot create duplicates.
- Learn it in: Webhooks, Error handling and retries, n8n in depth, RPA basics, Message queues and background work
- Related: Retry with backoff, Webhook
- Identity and access management (also: IAM)
- The cloud system that controls who and what (people, services, machines) can do which actions on which resources.
- Learn it in: Identity and access, One cloud in depth, Cloud security basics
- Related: Least privilege
- Image (also: Docker image, container image)
- The read-only template a container starts from, built from a Dockerfile: a base system, your dependencies and your code.
- Learn it in: CI/CD with GitHub Actions, Docker and Compose
- Related: Container, Dockerfile
- Incident response
- The planned way a team detects, contains, fixes and learns from a failure in a live system, with clear roles and communication.
- Learn it in: Incident response basics, Phishing and social engineering, Incident response for AI
- Related: Runbook, Kill switch
- Index (also: database index)
- A structure a database keeps alongside a table so it can find rows by a column quickly instead of scanning every row.
- Learn it in: SQL for engineers, Databases: indexes, replication and sharding
- Related: Join
- India AI Governance Guidelines
- Guidelines released by MeitY in November 2025 setting principles for safe, trusted AI in India. They create no new AI law and rely on existing laws such as the DPDP Act and on sector regulators.
- Learn it in: AI rules in India and abroad, Writing an AI policy, AI risk frameworks
- Related: DPDP Act, EU AI Act, Responsible AI
- Indicator of compromise (IOC)
- A piece of evidence that a system may have been attacked, such as a known bad IP address, domain, file hash or email sender. IOCs are shared between teams to block and hunt.
- Learn it in: Incident response basics, Phishing and social engineering, Threat detection with MITRE ATT&CK, LLMs in the SOC
- Related: MITRE ATT&CK, Detection rule, Incident response
- Inference
- Running a trained model to get an output. Every time you send a prompt and get a reply, that is one inference call.
- Learn it in: Serving a model
- Related: Model, Latency
- Infrastructure as code (also: IaC, Terraform)
- Describing servers, networks, databases and permissions in files that a tool such as Terraform applies, so infrastructure is reviewed, versioned and reproducible.
- Learn it in: Environments, secrets and Terraform
- Related: CI/CD
- Insecure output handling
- Passing model output to a browser, database, shell or other system without checking it, so injected text becomes cross-site scripting, SQL or a command. The 2025 OWASP list calls it improper output handling.
- Learn it in: Threats to LLM apps: the OWASP Top 10
- Related: Prompt injection, OWASP Top 10 for LLM Applications, Guardrails
- ISO/IEC 42001
- An international standard for an AI management system: policies, roles, risk assessment and controls an organisation can be certified against.
- Learn it in: AI risk frameworks
- Related: NIST AI RMF, EU AI Act
J
- Jailbreak
- A prompt, often role-play or persuasion, written to make a model ignore its safety rules and produce output it is meant to refuse.
- Learn it in: Red-team an AI system, Security tests in CI for AI apps
- Related: Prompt injection, Guardrails, Red teaming
- Join (also: SQL join, inner join, left join)
- An SQL operation that combines rows from two tables where a column matches, such as orders with their customers.
- Learn it in: SQL for engineers
- Related: Window function
- JSON
- A text format for structured data made of objects with named fields and lists, used by most web APIs.
- Learn it in: Lists and dictionaries, Call a web API and read JSON
- Related: API, Dictionary
- JSON Schema
- A standard way to describe the shape of JSON data: which fields exist, their types and which are required.
- Learn it in: Reading and shaping JSON, Get structured JSON output, Designing tools for agents, Tool calling
- Related: Structured output, Tool calling
K
- Keyword search (also: BM25, lexical search)
- Search that matches the exact words in the query, usually ranked with a formula such as BM25. Strong on names, codes and rare terms.
- Learn it in: Hybrid search, reranking and filters
- Related: Semantic search, Hybrid search
- Kill switch
- A quick, reliable way to stop an AI system or agent at once, used when it misbehaves or costs run away.
- Learn it in: Loop limits, budgets and kill switches, Incident response for AI
- Related: Guardrails, Budget limit
- Knowledge graph
- Data stored as entities (people, companies, products) and the relationships between them, so questions can follow links.
- Learn it in: Graph RAG
- Related: Graph RAG
- KPI (also: key performance indicator)
- Key performance indicator: the one or few numbers that show whether a project achieved its goal, measured against a baseline.
- Learn it in: Mapping a process, Choosing a problem for AI, Designing human in the loop, Measuring return on investment, AI PM interviews, Choosing a metric, Change management and training, Discovery, Demos, risk logs and escalation, Talking to executives, A case study, end to end, The FDE interview loop, The KPI that matters, Proof of concept, pilot, production, Portfolio and job search, Writing a scope of work
- Related: Baseline, Discovery
- Kubernetes (also: K8s)
- A system that runs containers across many machines, restarting failed ones, scaling them up and down and routing traffic to them.
- Learn it in: Kubernetes and Helm basics
- Related: Container, Helm, Pod
- KV cache
- Memory a model keeps during generation for every token in the context, so it need not recompute it. It grows with context length and concurrent requests.
- Learn it in: GPUs and quantization, Inside LLM inference: prefill, decode and the KV cache
- Related: GPU, vLLM
L
- Label (also: target)
- The correct answer for a training example, such as whether a loan was repaid, that a supervised model learns to predict.
- Related: Feature, Supervised learning
- LangGraph
- An open-source framework for building agents as graphs of steps with shared state, checkpoints and pauses for human approval.
- Learn it in: LangGraph: agents as graphs
- Related: Agent, Checkpoint
- Large language model (also: LLM)
- A model trained on huge amounts of text to predict the next token. That one skill lets it write, summarise, translate, answer questions and write code.
- It does not look facts up. It produces the most likely continuation of the text it was given, which is why it can sound confident and still be wrong.
- Learn it in: AI careers in India, A short history of AI, What AI cannot do, What AI, machine learning and a language model are, How AI products work
- Related: Token, Next-token prediction, Hallucination
- Latency
- How long a user waits for a response. For chat, the time to the first token matters most.
- Learn it in: Serving an AI app, Serving a model, How to approach system design, Caching and speed, Deploy and monitor an LLM app, The design interview, Quality, cost and speed trade-offs
- Related: Streaming, Caching, Latency budget
- Latency budget
- The longest response time you allow, split across the steps of a request, such as retrieval, model call and post-processing.
- Learn it in: Latency budgets and load testing
- Related: Latency
- Layout analysis
- Detecting a page's structure (headings, columns, tables, figures, reading order) so text is extracted in the right order and tables stay intact.
- Learn it in: Parsing documents and OCR
- Related: Document parsing, OCR
- Leaderboard
- A public ranking of models by their scores on a chosen set of benchmarks, or by people voting between two answers. Which tests it uses decides who comes first.
- Learn it in: Choosing a model: benchmarks, leaderboards and your own tests
- Related: Benchmark, Evals
- Learning rate
- How big a step training takes when it adjusts a model's weights. Too high and training is unstable; too low and it learns slowly or not at all.
- Learn it in: Fine-tuning in practice: tools, settings and reading the loss curve
- Related: Epoch, Batch size, Loss function
- Least privilege
- Giving a person, service or agent only the permissions it needs for its task, and nothing more.
- Learn it in: Linux for security, Identity and access, Cloud security basics, Threats to LLM apps: the OWASP Top 10, MCP servers and clients, Permissions and sandboxing, Securing agent tools and MCP servers, Threat model an AI system, Defend against prompt injection, Security questionnaires, Security and privacy
- Related: Sandbox, Guardrails
- Lexical similarity (also: BLEU, ROUGE, n-gram overlap)
- How many words or word sequences an answer shares with a reference answer. BLEU and ROUGE are common measures. It rewards matching words, not matching meaning.
- Learn it in: Scoring open-ended answers: exact checks, similarity and side-by-side comparison
- Related: Semantic similarity, Exact match
- Linked list (also: singly linked list, doubly linked list)
- A chain of nodes where each node holds a value and a pointer to the next node, so adding or removing at a known point needs no shifting.
- Undo history and music playlists can be linked lists: you insert or remove a step by changing two pointers, not by moving every item.
- Learn it in: Stacks, queues and linked lists
- Related: Array, Stack
- Linux permissions (also: chmod, file permissions)
- Read, write and execute rights on a Linux file or folder, set separately for the owner, the group and everyone else, and changed with chmod and chown.
- Learn it in: Linux for security
- Related: Least privilege, SSH
- List
- An ordered collection of values in Python, written in square brackets, such as marks = [80, 72, 91]. Items are reached by position.
- Learn it in: Variables, loops and functions in Python, Lists and dictionaries
- Related: Dictionary
- LLM as judge (also: LLM judge)
- Using a model to grade another model's outputs against a rubric. Fast and cheap, but it must be checked against human grades.
- Learn it in: Test sets and evals, Evaluating agents, LLM judges and their limits, Online evals and feedback
- Related: Evals
- LLM gateway (also: AI gateway, model gateway)
- A single service in front of all your model providers that handles routing, fallbacks, rate limits, keys and cost tracking.
- Learn it in: LLM gateways, Grow an AI app in layers: context, guardrails, routing, caching and agents, Batching and model tiering
- Related: Model tiering, Rate limit
- Load balancer (also: LB, reverse proxy)
- A component that spreads incoming requests across several servers and stops sending traffic to servers that are unhealthy.
- Like a supervisor at a bank directing each customer to the next free counter, a load balancer sends each request to a server that can take it.
- Learn it in: Scaling and load balancing
- Related: Horizontal scaling, Throughput
- Load testing
- Sending realistic traffic at a system, rising step by step, to see how response times and errors change before real users find the limit.
- Learn it in: Latency budgets and load testing
- Related: 95th percentile latency, Latency budget
- Logging
- Recording what a program does as it runs, with a level (debug, info, warning, error) and a time, so you can understand problems later.
- Learn it in: Project tooling: environments, tests and logs, Logs and SIEM, Audit logs, Cloud security basics, Logs and traces for agents
- Related: Observability, Tracing
- Logits
- The raw scores a model gives every possible next token before they are turned into chances. Higher means more likely. Temperature changes the logits before this step.
- Learn it in: How a model picks its next word: temperature, top-p and top-k
- Related: Sampling, Temperature, Next-token prediction
- Loop (also: for loop, while loop)
- Code that repeats a block of steps, for example once for every item in a list.
- Learn it in: Variables, loops and functions in Python
- Related: List
- LoRA (low-rank adaptation)
- A way to fine-tune a model by freezing its weights and training two small extra matrices per layer instead. It trains well under 1% of the parameters, so it needs far less memory than full fine-tuning.
- Learn it in: LoRA, QLoRA and the GPU memory maths, Model merging: combining fine-tuned models into one
- Related: Fine-tuning, QLoRA, Weights
- Loss function (also: cost function)
- A formula that measures how wrong a model's predictions are on training data. Training tries to make it as small as possible.
- Learn it in: Regression
- Related: Gradient descent
- Lost in the middle
- The finding that language models use information near the start and end of a long context better than information in the middle.
- Learn it in: Long context or RAG
- Related: Context window, Retrieval-augmented generation
- Low-resource language
- A language with little text available online for training AI models. Many Indian languages, such as Odia, Assamese and Punjabi, are low-resource despite having millions of speakers.
- Learn it in: AI in Indian languages: tokens, quality gaps and what to use
- Related: Token fertility, Tokenizer
M
- Machine learning (also: ML)
- A way of building software that learns patterns from examples instead of following hand-written rules.
- Show a machine learning system thousands of past loan applications and whether each was repaid, and it learns which patterns predict repayment. It can then score a new application it has never seen.
- Learn it in: Bias and fairness, Choosing your path, A short history of AI, What AI, machine learning and a language model are
- Related: Artificial intelligence, Training, Model
- Matrix
- A rectangular table of numbers. A dataset is a matrix with rows for examples and columns for features; model weights are often matrices too.
- Learn it in: Maths for machine learning, NumPy and plotting
- Related: Vector
- MCP server
- A small program that exposes tools, data or prompts through the Model Context Protocol so any MCP-capable AI app can use them.
- Learn it in: MCP servers and clients
- Related: Model Context Protocol, Tool calling
- Mean and median
- Two kinds of average. The mean adds values and divides by the count; the median is the middle value. With a few extreme values, such as very high incomes, the median is more typical.
- Learn it in: Statistics for machine learning
- Related: Probability distribution
- Membership inference (also: membership inference attack)
- An attack that works out whether a particular record, such as one person's data, was in a model's training set, often because the model is more confident on data it has seen.
- Learn it in: Model extraction and privacy attacks
- Related: Model extraction, Training-data extraction, Personal data
- Memoization (also: memoisation, caching results)
- Storing the result of a function call so that the next call with the same input returns the saved answer instead of computing it again.
- A naive recursive Fibonacci recomputes the same values millions of times. With memoization each value is computed once, turning hours into milliseconds.
- Learn it in: Recursion and dynamic programming
- Related: Dynamic programming, Caching
- Merge conflict
- What happens when two branches change the same lines differently. Git stops and asks a person to choose the right result.
- Learn it in: Git branches, merging and rebasing, Git with a team
- Related: Branch
- Message queue (also: Kafka, RabbitMQ, SQS, pub/sub)
- A service that holds messages from one part of a system until another part is ready to process them, so slow work can happen in the background.
- When you place an order, the shop saves it and puts an 'order placed' message on a queue. Separate workers then send the SMS, update stock and email the invoice.
- Learn it in: Message queues and background work, Design walkthrough: a URL shortener
- Related: Queue, Idempotency
- Metadata filter
- Restricting a search to documents with certain properties, such as a date range, department or document type, before or alongside the similarity search.
- Learn it in: Hybrid search, reranking and filters, Metadata that helps retrieval, Permission-aware RAG
- Related: Hybrid search, Vector database
- Misinformation (also: disinformation)
- False or misleading information spread whether or not the person sharing it knows it is false. Disinformation is the deliberate kind, spread to deceive.
- Learn it in: Checking an AI answer, Deepfakes and misinformation
- Related: Deepfake, Hallucination
- MITRE ATLAS
- A free public knowledge base from MITRE of the tactics and techniques attackers use against AI and machine learning systems, with real case studies, laid out like MITRE ATT&CK.
- Learn it in: Threat model an AI system, Data poisoning, backdoors and adversarial examples
- Related: Threat modelling, Red teaming, Prompt injection
- MITRE ATT&CK
- A free, public knowledge base from MITRE of how real attackers behave, organised as tactics (the goal, such as credential access) and techniques (the method, such as brute force), each with an id like T1110.
- Learn it in: Phishing and social engineering, Threat detection with MITRE ATT&CK
- Related: Detection rule, Sigma rule, Indicator of compromise (IOC)
- Mixture of experts (MoE) (also: MoE)
- A model split into several expert sub-networks, where only a few run for each token. It has many parameters in total but the speed and cost of a much smaller model.
- Learn it in: Model size and scaling laws: parameters, training data and compute
- Related: Parameters
- Mocking
- Replacing a slow or external part, such as an API or model call, with a fake that returns fixed answers, so a test is fast and repeatable.
- Learn it in: Writing tests with pytest
- Related: Unit test, Dependency injection
- Model
- The learned part of an AI system: a file of numbers (weights) that turns an input into an output, such as text into a reply.
- Related: Weights, Training, Inference
- Model card
- A short document published with a model that says who made it, what data it was trained on, what it is for, how it was evaluated, its limits and its licence.
- Learn it in: Securing the AI supply chain
- Related: AI supply chain, AI bill of materials, Open-weight model
- Model collapse
- The decline seen when models are trained again and again on data made by models. Rare cases fade and errors pile up, so outputs grow bland and less accurate.
- Learn it in: Synthetic data and distillation: training with model-made data
- Related: Synthetic data, Distillation
- Model Context Protocol (also: MCP)
- An open standard for connecting AI apps to tools and data. A tool is wrapped once as an MCP server and any MCP-capable app can use it.
- Learn it in: MCP servers and clients, Securing agent tools and MCP servers
- Related: Tool calling, Agent
- Model extraction (also: model stealing, model theft)
- Copying a model by sending it many queries through its API and training a new model on the answers, without access to the original weights.
- Learn it in: Model extraction and privacy attacks
- Related: Membership inference, Training-data extraction, Rate limit
- Model merging (also: merging)
- Combining the weights of two or more models, usually fine-tuned from the same base, into one model, without serving them separately or training from scratch.
- Learn it in: Model merging: combining fine-tuned models into one
- Related: Fine-tuning, LoRA (low-rank adaptation), Task vector
- Model registry
- A store of trained model versions with their metadata and stage (staging, production), used to promote and roll back models.
- Learn it in: Data versioning, Experiment tracking and model registry, Serving a model
- Related: Experiment tracking
- Model router (also: intent classifier)
- A small, fast step that decides where each request goes: which model, which tool, a stock reply or a person, based on what the user is trying to do.
- Learn it in: Grow an AI app in layers: context, guardrails, routing, caching and agents
- Related: LLM gateway, Guardrails
- Model tiering (also: model routing)
- Sending easy requests to a small, cheap model and only hard ones to a large, expensive model.
- Learn it in: Pricing AI features, Cost of ownership, Batching and model tiering, Cost control, Quality, cost and speed trade-offs
- Related: Cost per request, LLM gateway
- Monitoring
- Watching a live system's key numbers, such as errors, latency, cost and quality, and alerting a person when one goes wrong.
- Learn it in: Error handling and retries, ML interviews, Monitoring and drift, Deploy and monitor an LLM app, Design an AI system
- Related: Observability, Drift
- Motion graphics
- Animated text, shapes, charts and icons used to explain an idea, as in explainer videos. Often made from code or templates rather than filmed.
- Learn it in: Make animated explainer videos with Claude
- Related: Text-to-video
- Multi-agent system
- A design where several agents, each with its own role, tools and context, work together on a task, usually coordinated by a lead agent.
- Learn it in: Multi-agent systems
- Related: Sub-agent, Agent
- Multi-factor authentication (also: MFA, 2FA, two-step verification)
- Signing in with two or more different kinds of proof, such as a password plus an authenticator app code or a passkey, so a stolen password alone is not enough.
- Learn it in: Identity and access, AI-powered attacks and how defenders respond
- Related: Authentication, OAuth 2.0
- Multimodal RAG
- Retrieval and answering over images, charts, tables and page layouts as well as text, often with a vision-capable model.
- Learn it in: Multimodal RAG
- Related: Retrieval-augmented generation, Vision-language model
N
- n8n
- A source-available ("fair-code") workflow automation tool with a visual editor, many app connectors and AI steps. It can be self-hosted or used as a cloud service.
- Learn it in: Make, Zapier and Power Automate, Your first n8n workflow, n8n in depth
- Related: Automation, Node
- Neural network
- A model made of layers of simple units, each combining inputs with weights and passing the result on, able to learn complex patterns.
- Learn it in: Neural networks
- Related: Backpropagation, Deep learning
- Next-token prediction
- What a language model actually does: given the text so far, it scores every possible next token and picks one, then repeats.
- Learn it in: Perplexity and cross-entropy: how language models are measured
- Related: Token, Sampling, Large language model
- NIST AI RMF
- The US National Institute of Standards and Technology's voluntary AI Risk Management Framework: govern, map, measure and manage AI risks.
- Learn it in: AI risk frameworks
- Related: ISO/IEC 42001, EU AI Act
- Node (also: step, module, action)
- One step in a visual automation workflow, such as reading an email, calling an API, transforming data or sending a message.
- Learn it in: What automation is, Your first n8n workflow, n8n in depth
- Related: Trigger, Automation
O
- OAuth 2.0
- A standard for letting an app act on a user's behalf with limited permission, using tokens instead of the user's password.
- Learn it in: Identity and access, Single sign-on
- Related: OpenID Connect, Single sign-on
- Observability
- Being able to tell what a live system is doing and why, from its logs, metrics and traces.
- Learn it in: Tracing LLM apps, Deploy and monitor an LLM app
- Related: Tracing, Monitoring
- OCR (also: optical character recognition)
- Optical character recognition: turning an image of text, such as a scanned page or a photo of a bill, into text a program can read.
- Learn it in: Files, tables and PDFs, Extracting data from documents, Parsing documents and OCR, Multimodal RAG
- Related: Document parsing
- Online evaluation
- Scoring a live system's real traffic, by automatic checks, model judges or user feedback, to catch quality problems after launch.
- Learn it in: Data strategy for AI products, Monitoring and drift, Online evals and feedback, Deploy and monitor an LLM app
- Related: Evals, Monitoring, Drift
- Open-weight model (also: open model)
- A model whose weights are published, so you can run it on your own machines instead of calling a vendor's API.
- Learn it in: Securing the AI supply chain, Choosing a model: benchmarks, leaderboards and your own tests, Open-weight models with vLLM
- Related: Weights, Self-hosting
- OpenID Connect (also: OIDC)
- A sign-in standard built on OAuth 2.0. It gives the app an ID token saying who the user is, plus claims such as email and groups.
- Learn it in: Identity and access, Single sign-on
- Related: OAuth 2.0, Single sign-on
- Overfitting
- When a model learns the training data too closely, including its noise, so it performs well on training data but poorly on new data.
- Learn it in: Cross-validation and leakage, Regression, Trees and gradient boosting, CNNs for images, Fine-tuning in practice: tools, settings and reading the loss curve, How fine-tuning works: data, training and what can go wrong, ML interviews
- Related: Cross-validation, Regularisation
- OWASP Top 10
- A widely used list, published by the OWASP Foundation, of the ten most critical security risk categories in web applications, updated every few years; the current edition is 2025.
- Learn it in: Web app security and the OWASP Top 10
- Related: SQL injection, Cross-site scripting
- OWASP Top 10 for LLM Applications
- A community list from the OWASP GenAI Security Project of the ten most important security risks in apps built on large language models, from prompt injection to unbounded consumption, with fixes for each. It is updated as attacks change.
- Learn it in: Threats to LLM apps: the OWASP Top 10
- Related: Prompt injection, Excessive agency, Insecure output handling
P
- Pairwise comparison (also: comparative evaluation, side-by-side evaluation)
- Showing a person or a judge model two answers side by side and asking which is better. Many such votes can be turned into a ranking of models.
- Learn it in: Scoring open-ended answers: exact checks, similarity and side-by-side comparison
- Related: LLM as judge, Evals
- Parameters (also: model size)
- The numbers inside a model that training adjusts, often counted in billions (the 8B in a model's name). More parameters can learn more, but need more memory and compute.
- Learn it in: Model size and scaling laws: parameters, training data and compute
- Related: Weights, Scaling law, Mixture of experts (MoE)
- Parent-child chunks (also: small-to-big retrieval, parent document retrieval)
- Searching small chunks for precision but handing the model the larger section each one came from, so it gets enough context to answer.
- Learn it in: Query rewriting, HyDE and parent-child chunks, Long documents without vectors
- Related: Chunking, Retrieval
- Permission-aware RAG (also: security trimming, document-level access control)
- Retrieval that only returns passages the current user is allowed to see, by filtering on access rights stored with each document.
- Learn it in: Metadata that helps retrieval, Permission-aware RAG, A case study, end to end
- Related: Retrieval-augmented generation, Role-based access control, Metadata filter
- Perplexity (also: PPL)
- How surprised a language model is by a text, on average: roughly how many choices it is weighing for each next token. Lower means the model predicts the text better.
- Learn it in: Perplexity and cross-entropy: how language models are measured
- Related: Cross-entropy, Next-token prediction
- Personal data (also: PII, personally identifiable information)
- Information that identifies a person, such as a name, phone number, Aadhaar or PAN number, address or email.
- Learn it in: AI rules in India and abroad, Your company's AI policy, Images, voice and other AI tools, Use AI safely at work, Data strategy for AI products, Chat assistants for websites and WhatsApp, Audit logs, Data quality and personal data, Agent memory, Guardrails, Model extraction and privacy attacks, Security and privacy
- Related: DPDP Act, Data exfiltration
- Phishing
- A message, usually email, SMS or a chat, that pretends to come from someone trusted to trick the reader into clicking a link, opening a file, entering a password or sending money.
- Learn it in: Phishing and social engineering
- Related: Social engineering, MITRE ATT&CK
- Pickle deserialisation (also: pickle deserialization, unsafe model loading)
- Loading a file in Python's pickle format, which can call functions while it loads. A malicious pickle model file can run an attacker's code; formats like safetensors avoid this.
- Learn it in: Securing the AI supply chain
- Related: AI supply chain, Open-weight model
- Pilot
- Running a system with a limited group of real users in daily work, to measure value and risks before full rollout.
- Learn it in: Proof of concept, pilot, production
- Related: Proof of concept, Stage gate
- pip
- Python's standard tool for installing packages from the Python Package Index.
- Learn it in: Packages and virtual environments
- Related: Virtual environment
- Pipeline parallelism
- Putting different groups of layers on different GPUs, so data flows from one GPU to the next. Good for training throughput; adds delay for single requests.
- Learn it in: Running models on many GPUs: replicas, tensor and pipeline parallelism
- Related: Tensor parallelism, Data parallelism
- Planning (also: plan-and-execute)
- An agent writing out the steps it intends to take before acting, so it can follow, check and revise them.
- Learn it in: Planning and reflection
- Related: Reflection, Agent loop
- Pod
- The smallest unit Kubernetes runs: one or more containers that share a network address and storage, started and stopped together.
- Learn it in: Kubernetes and Helm basics
- Related: Kubernetes, Container
- Port
- A number from 0 to 65535 that tells a machine which service a network connection is for, such as 22 for SSH or 443 for HTTPS.
- Learn it in: Linux for security, Networking for security
- Related: TCP/IP, Firewall
- Pre-training (also: pretraining)
- The first and most expensive stage of making a language model: training it on a huge amount of text to predict the next token. The result is a base model.
- Learn it in: How a chat model is made: pre-training, SFT and preference tuning
- Related: Base model, Next-token prediction, Training
- Precision and recall
- Precision: of the cases flagged, how many were truly positive. Recall: of the actual positive cases, how many were flagged. Raising one often lowers the other.
- Learn it in: Choosing a metric, Classification, Machine learning for threat detection, ML interviews
- Related: Confusion matrix, Classification
- Prefill (also: prefill and decode)
- The first stage of generating a reply: the model reads the whole prompt at once and fills its KV cache. Long prompts make it slower, which delays the first token.
- Learn it in: Inside LLM inference: prefill, decode and the KV cache
- Related: KV cache, Time to first token (TTFT)
- Probability distribution
- A description of how likely each possible value is, such as the spread of delivery times or loan amounts.
- Learn it in: Maths for machine learning, Statistics for machine learning
- Related: Mean and median
- Process map
- A diagram of the steps in a business process, who does each, the systems used and where delays or errors happen.
- Learn it in: Mapping a process, Scoping an automation project
- Related: Automation, Discovery
- Progress file (also: Progress log, Notes file)
- A file on disk where an agent records what it has done, what is left and what to do next, so a new session with an empty context can pick up the work.
- Anthropic's long-running agent harness pairs a progress log with a JSON feature list and git commits. Each new session reads them first, the way a new shift reads the handover notes.
- Learn it in: Advanced harness engineering, Harnesses for long-running agents
- Related: Agent memory, Context compaction, Handover
- Prompt
- Everything you send to a model in one call: instructions, examples, documents and the question.
- Learn it in: AI for writing and summarising, Prompting without jargon, Use a chat assistant for a real task and check the answer, Call an LLM API from code
- Related: System prompt, Few-shot prompting, Prompt engineering
- Prompt engineering
- Designing and testing prompts so a model gives reliable, useful output for a task.
- In practice it means clear instructions, the right context, a worked example and a fixed output format, then checking the results on real cases and adjusting.
- Learn it in: AI image generation and editing with Gemini (Nano Banana) and Midjourney, Vibe coding with Claude Code: build real apps by describing them, Make animated explainer videos with Claude, System prompts and prompt design, Prompting, RAG or fine-tuning
- Related: Prompt, Few-shot prompting, Evals
- Prompt injection
- An attack where text from a user, web page or document contains instructions that trick the model into ignoring its own rules.
- Learn it in: LLMs in the SOC, Threats to LLM apps: the OWASP Top 10, Permissions and sandboxing, Threat model an AI system, Browser and computer-use agents, Guardrails, Defend against prompt injection, Red-team an AI system, Security and privacy, Security tests in CI for AI apps
- Related: Guardrails, System prompt, Data exfiltration
- Prompt registry
- A store of versioned prompts with names and aliases, so the prompt in production is known, reviewed and can be rolled back.
- Learn it in: Versioning prompts and models
- Related: Regression test, CI/CD
- Proof of concept (also: POC)
- A small, time-boxed build that tests whether an idea can work on the client's real data, before investing in a pilot.
- Learn it in: Feasibility and build versus buy, Proof of concept, pilot, production
- Related: Pilot, Stage gate
- Pull request (also: PR, merge request)
- A request to merge a branch into the main line, where teammates review the changes, comment and run automated checks first.
- Learn it in: Git with a team, Code review: giving and getting good reviews
- Related: Branch, Code review
- Pydantic
- A Python library that checks data against a model you define with type hints, converting and rejecting bad values with clear errors.
- Learn it in: Python for production, Get structured JSON output
- Related: Type hint, JSON Schema
- PyRIT
- Microsoft's open-source Python Risk Identification Tool for red teaming generative AI: it sends attack prompts to a target, scores the replies and keeps a record of every attempt.
- Learn it in: Red-team an AI system
- Related: Red teaming, garak
- pytest
- The most popular Python testing tool. You write plain functions starting with test_ and use assert; pytest finds and runs them.
- Learn it in: Project tooling: environments, tests and logs, Writing tests with pytest
- Related: Unit test, Fixture
- Python
- A popular, readable programming language widely used for data work, automation and AI. Most AI tools and libraries offer Python first.
- Learn it in: Set up Python and run your first file, Packages and virtual environments, Tables with pandas
- Related: Variable, Function
Q
- QLoRA
- LoRA applied to a model whose frozen weights are stored in 4-bit form. It cuts memory enough to fine-tune a large model on a single consumer or cloud GPU.
- Learn it in: LoRA, QLoRA and the GPU memory maths
- Related: LoRA (low-rank adaptation), Quantization, Fine-tuning
- Quantization
- Storing a model's weights with fewer bits, such as 8 or 4 instead of 16, so it needs less memory and runs faster, at some cost in quality.
- Learn it in: GPUs and quantization, LoRA, QLoRA and the GPU memory maths, Open-weight models with vLLM
- Related: vLLM, Open-weight model
- Query rewriting (also: query expansion, multi-query)
- Using a model to turn a user's question into one or more better search queries before retrieval, such as expanding abbreviations or splitting a two-part question.
- Learn it in: Query rewriting, HyDE and parent-child chunks
- Related: Retrieval, HyDE
- Queue (also: FIFO)
- A collection where the first item added is the first one taken out, like a line at a ticket counter.
- A print queue and a breadth-first search both use a queue so items are handled in the order they arrived.
- Learn it in: Stacks, queues and linked lists
- Related: Stack, Message queue, Breadth-first search
R
- Random forest
- Many decision trees trained on random parts of the data and features, whose predictions are averaged. Robust and easy to use.
- Learn it in: Trees and gradient boosting
- Related: Decision tree, Gradient boosting
- Rate limit
- A cap on how many requests or tokens you may send in a period. Going over returns an error (often 429) until the window resets.
- Learn it in: Error handling and retries, n8n in depth, Errors, retries and cost, Enterprise connectors, Loop limits, budgets and kill switches, LLM gateways, Model extraction and privacy attacks, Latency budgets and load testing
- Related: Retry with backoff, HTTP status code
- ReAct
- A common agent pattern that alternates reasoning ("what should I do next?") with acting (calling a tool) and observing the result.
- Learn it in: Planning and reflection
- Related: Agent loop, Step-by-step reasoning
- Reasoning model
- A model trained to think through a problem in hidden or visible steps before it answers. Better on hard problems; slower and costlier.
- Learn it in: Reasoning models and test-time compute: when to let the model think longer
- Related: Step-by-step reasoning, Latency
- Rebase
- Replaying your branch's commits on top of another branch, so history reads as a straight line. Never rebase commits other people have already pulled.
- Learn it in: Git branches, merging and rebasing
- Related: Branch, Merge conflict
- Recall at k (also: recall@k)
- Of the passages that should have been found, the share that appear in the top k results. The first number to check in a RAG system.
- Learn it in: Check whether retrieval works
- Related: Retrieval, Golden set
- Reciprocal rank fusion (also: RRF)
- A simple way to merge several ranked result lists by giving each item points based on its rank in each list, then sorting by total points.
- Learn it in: Hybrid search, reranking and filters
- Related: Hybrid search
- Recursion (also: recursive function)
- When a function solves a problem by calling itself on a smaller version of the same problem, until it reaches a base case it can answer directly.
- Counting files in a folder is recursive: count the files here, then call the same function on each sub-folder. Without a base case it never stops.
- Learn it in: Recursion and dynamic programming
- Related: Dynamic programming, Stack, Depth-first search
- Red teaming
- Authorised testing where people and tools play the attacker against your own AI system, trying jailbreaks, prompt injection and data leaks, so weaknesses are found and fixed before real attackers find them.
- Learn it in: Red-team an AI system
- Related: Jailbreak, Prompt injection, PyRIT, garak
- Refactoring
- Changing how code is written without changing what it does, to make it easier to read, test and change. Tests make it safe.
- Learn it in: Clean code and refactoring, Design principles: modules, SOLID and dependency injection
- Related: Code smell, Unit test
- Reflection (also: self-critique)
- An agent checking its own output or progress and correcting course, for example by critiquing a draft before sending it.
- Learn it in: Planning and reflection
- Related: Planning
- Region (also: cloud region)
- A geographic area where a cloud provider runs data centres, such as Mumbai or Hyderabad. Choosing one affects latency, price, available models and where data legally sits.
- Learn it in: One cloud in depth
- Related: Data residency
- Regression
- Predicting a number, such as a house price or delivery time.
- Learn it in: Regression
- Related: Classification, Supervised learning
- Regression test
- A test that reruns old cases after a change to catch anything that used to work and now does not.
- Learn it in: CI/CD with GitHub Actions, Experiment tracking and model registry, Versioning prompts and models, Test sets and evals, Regression tests for prompts, Security tests in CI for AI apps
- Related: Evals, Golden set
- Regularisation (also: regularization)
- Techniques that discourage a model from becoming too complex, such as penalising large weights, to reduce overfitting.
- Related: Overfitting
- Replication (also: read replica, primary and replica)
- Keeping copies of the same data on more than one database server, so reads can be spread out and the data survives if one server fails.
- A primary database takes all writes and copies them to two replicas that serve reads. If the primary fails, a replica can take over.
- Learn it in: Databases: indexes, replication and sharding
- Related: Sharding, Index
- Reranking (also: reranker)
- A second, more careful pass that reorders the top search results by how well each actually answers the query.
- Learn it in: Hybrid search, reranking and filters, Vector search
- Related: Hybrid search, Retrieval, Cross-encoder
- Responsible AI
- Building and using AI so it is fair, safe, private, explainable and someone is accountable for it.
- Learn it in: Bias and fairness, Your company's AI policy, Use AI safely at work, Writing an AI policy, Accountability for agent actions, AI risk frameworks
- Related: Bias, EU AI Act, Human in the loop
- REST API
- An API built on plain HTTP: resources at URLs, and methods such as GET to read and POST to create.
- Learn it in: Call a web API and read JSON, REST APIs with FastAPI
- Related: API, HTTP status code
- Retrieval
- Finding the documents or passages most relevant to a question before the model answers.
- Learn it in: Check whether retrieval works, Chunk documents for retrieval, Build a RAG app
- Related: Retrieval-augmented generation, Semantic search, Recall at k
- Retrieval-augmented generation (also: RAG)
- Answering a question by first retrieving relevant passages from your own documents and putting them in the prompt, so the model answers from them.
- RAG is how most companies put AI on their own data without retraining a model. Its quality depends mostly on retrieval: if the right passage is not found, the model cannot use it.
- Learn it in: AI careers in India, AI for research and analysis, Choosing your path, How AI products work, Chat assistants for websites and WhatsApp, Long context or RAG, Build a RAG app, Data poisoning, backdoors and adversarial examples, Graph RAG, Design an AI system, The design interview, Prompting, RAG or fine-tuning
- Related: Retrieval, Grounding, Chunking, Embedding
- Retry with backoff (also: exponential backoff)
- Trying a failed request again after waiting, and waiting longer after each failure, so a busy service can recover.
- Learn it in: Error handling and retries, Errors, retries and cost
- Related: Rate limit
- Return on investment (also: ROI)
- The value an investment returns compared with its cost, often shown as a percentage or a payback period.
- Learn it in: Measuring return on investment
- Related: Total cost of ownership, KPI
- Reward model
- A model trained on human comparisons of two answers to predict which answer people would prefer. RLHF uses its scores to tune a chat model.
- Learn it in: How a chat model is made: pre-training, SFT and preference tuning
- Related: RLHF (reinforcement learning from human feedback), DPO (direct preference optimisation)
- Risk log (also: risk register)
- A living list of what could go wrong, how likely and serious each risk is, who owns it and what is being done about it.
- Learn it in: Demos, risk logs and escalation
- Related: Escalation
- RLHF (reinforcement learning from human feedback)
- A way to tune a model toward answers people prefer: collect comparisons of two answers, train a reward model on them, then train the chat model to score well with that reward model.
- Learn it in: How a chat model is made: pre-training, SFT and preference tuning
- Related: Reward model, DPO (direct preference optimisation), Supervised fine-tuning (SFT)
- Robotic process automation (also: RPA)
- Software robots that operate applications through their user interfaces, clicking and typing like a person, for systems without APIs.
- Learn it in: Make, Zapier and Power Automate, RPA basics
- Related: Automation, Computer-use agent
- ROC AUC (also: AUC)
- A score up to 1 (perfect) for how well a classifier ranks positives above negatives across all thresholds. 0.5 means no better than random guessing.
- Learn it in: Choosing a metric, Classification
- Related: Precision and recall
- Role-based access control (also: RBAC)
- Granting permissions to roles (such as "claims officer") and roles to people, so access follows a person's job.
- Learn it in: Identity and access, Permission-aware RAG
- Related: Least privilege, Permission-aware RAG
- Runbook
- Step-by-step instructions for operating a system and handling known problems, written for the people on call.
- Learn it in: Incident response basics, Incident response for AI, Documentation and handover
- Related: Handover
S
- SAML (also: SAML 2.0)
- An older XML-based single sign-on standard, still common in large enterprises, where the company's identity provider vouches for the user to the app.
- Learn it in: Single sign-on
- Related: Single sign-on, OpenID Connect
- Sampling
- How a model picks the next token from its scored options. Settings like temperature change how adventurous the pick is.
- Learn it in: How a model picks its next word: temperature, top-p and top-k
- Related: Temperature, Next-token prediction
- Sandbox
- An isolated environment where untrusted code, such as code an agent writes, runs without access to anything it could damage.
- Learn it in: Permissions and sandboxing, Browser and computer-use agents, Advanced harness engineering
- Related: Least privilege
- Scaling law (also: scaling laws)
- A rule, found by experiment, for how a model's quality improves as you add parameters, training data and compute. It lets labs predict a large model's results from small ones.
- Learn it in: Model size and scaling laws: parameters, training data and compute
- Related: Compute-optimal, Parameters
- Scope creep
- The gradual growth of a project's requirements beyond what was agreed, without matching changes to time, cost or other scope.
- Learn it in: Scoping an automation project, Managing scope creep, Writing a scope of work
- Related: Scope of work, Change request
- Scope of work (also: SOW, statement of work)
- A document agreeing what will be delivered, what will not, by when, how success is judged and what each side must provide.
- Learn it in: Scoping an automation project, The FDE interview loop, Managing scope creep, Writing a scope of work
- Related: Acceptance criteria, Scope creep
- Secrets management
- Keeping passwords, keys and tokens out of code and logs, storing them in a secure place and giving each system only the ones it needs.
- Learn it in: Cryptography basics, Linux, the shell and SSH, Cloud security basics, Environments, secrets and Terraform, Security and privacy, Security tests in CI for AI apps
- Related: Environment variable
- Security copilot
- An AI assistant, usually built on a large language model, that helps security analysts summarise alerts, explain logs, write queries and draft reports. Its output is a draft for a person to check, not a verdict.
- Learn it in: LLMs in the SOC
- Related: Large language model, Hallucination, Human in the loop
- Security questionnaire (also: vendor risk assessment)
- A long list of questions a client sends a vendor about its security, privacy and resilience before approving it.
- Learn it in: Security questionnaires, Run an AI security programme
- Related: SOC 2
- Self-hosting
- Running a model or service on machines you control, often because data may not leave the company.
- Learn it in: Deployment models, Cost of ownership, Open-weight models with vLLM
- Related: Open-weight model
- Semantic cache
- A cache that reuses an earlier answer when a new question means the same thing, not just when it is word-for-word identical. It saves cost but can return wrong answers if the match is loose.
- Learn it in: Grow an AI app in layers: context, guardrails, routing, caching and agents
- Related: Cache invalidation, KV cache, Semantic similarity
- Semantic search (also: vector search)
- Search that matches by meaning, using embeddings, rather than by shared keywords.
- Learn it in: Embeddings, Vector search
- Related: Embedding, Keyword search, Hybrid search
- Semantic similarity (also: embedding similarity)
- How close two texts are in meaning, usually measured by turning both into embeddings and comparing them with cosine similarity.
- Learn it in: Scoring open-ended answers: exact checks, similarity and side-by-side comparison
- Related: Cosine similarity, Embedding, Lexical similarity
- Shadow AI
- AI tools that staff use for work without the company's approval or knowledge, such as pasting customer data into a personal chatbot account.
- Learn it in: Run an AI security programme
- Related: AI inventory, Data exfiltration
- SIEM
- Security information and event management: a system that collects logs from many sources into one place, searches them, links related events and raises alerts for analysts. Examples include Microsoft Sentinel, Splunk, Elastic Security and the open-source Wazuh.
- Learn it in: Incident response basics, Logs and SIEM, Threat detection with MITRE ATT&CK, Machine learning for threat detection
- Related: Logging, SOC (security operations centre), Alert fatigue
- Sigma rule
- A detection rule written in Sigma, an open YAML format that describes what to look for in logs without tying it to one product, so tools can convert it into Splunk, Sentinel, Elastic or other queries.
- Learn it in: Threat detection with MITRE ATT&CK
- Related: Detection rule, SIEM, MITRE ATT&CK
- Single sign-on (also: SSO)
- Signing in once with a company account and getting access to many apps, so each app does not keep its own passwords.
- Learn it in: Single sign-on, Permission-aware RAG
- Related: OAuth 2.0, OpenID Connect, SAML
- SOC (security operations centre)
- The team, and often the room, that watches an organisation's systems around the clock, triages security alerts, investigates suspicious activity and starts incident response.
- Learn it in: Incident response basics, Logs and SIEM, LLMs in the SOC
- Related: SIEM, Incident response, Alert fatigue
- SOC 2
- An independent audit report on a service provider's controls for security and, optionally, availability, confidentiality, processing integrity and privacy.
- Learn it in: Compliance for AI systems, Security questionnaires
- Related: Security questionnaire
- SOLID principles
- Five design guidelines: single responsibility, open for extension, substitutable subtypes, small interfaces, and depending on abstractions rather than concrete details.
- Learn it in: Design principles: modules, SOLID and dependency injection
- Related: Dependency injection, Refactoring
- Span
- One timed step inside a trace, such as a retrieval, a model call or a tool call, with its inputs, outputs and duration.
- Learn it in: Tracing LLM apps, Logs and traces for agents
- Related: Tracing
- Speculative decoding (also: assisted generation)
- A speed-up where a small, fast model drafts several tokens and the main model checks them all at once, keeping the ones it agrees with. The answer is the same, only faster.
- Learn it in: Inside LLM inference: prefill, decode and the KV cache
- Related: KV cache, Latency
- Speech-to-text
- Turning spoken audio into written text, also called transcription. The first step of every voice agent.
- Learn it in: Build AI voice agents with Vapi: phone and web assistants that talk
- Related: Voice agent, Text-to-speech
- SQL injection (also: SQLi)
- A flaw where user input is pasted into a database query, so crafted input changes the query itself. The fix is parameterised queries.
- Learn it in: Web app security and the OWASP Top 10
- Related: OWASP Top 10, Cross-site scripting
- SSH
- Secure Shell: an encrypted way to log in to and run commands on a remote machine, usually with a key pair instead of a password.
- Learn it in: Linux for security, Linux, the shell and SSH
- Related: Terminal
- Stack (also: LIFO)
- A collection where the last item added is the first one taken out, like a pile of plates.
- Your browser's back button is a stack of pages: the page you opened last is the first one you go back to.
- Learn it in: Stacks, queues and linked lists
- Related: Queue, Recursion
- Stage gate (also: go/no-go)
- An agreed checkpoint between project phases where evidence is reviewed and a decision is made to continue, change or stop.
- Learn it in: Demos, risk logs and escalation, Talking to executives, A case study, end to end, Managing scope creep, Proof of concept, pilot, production
- Related: Proof of concept, Pilot
- STAR method (also: STAR)
- A way to answer behavioural interview questions: Situation, Task, Action, Result.
- Learn it in: The FDE interview loop, Portfolio and job search
- Step-by-step reasoning (also: chain of thought)
- Asking a model to work through a problem in steps before answering, which often improves accuracy on multi-step tasks.
- Learn it in: Reasoning models and test-time compute: when to let the model think longer
- Related: Reasoning model, Prompt engineering
- Streaming
- Sending a model's reply piece by piece as it is written, so the user sees text appear instead of waiting for the whole answer.
- Learn it in: Serving an AI app, Caching and speed
- Related: Latency
- STRIDE
- A checklist of six threat types used in threat modelling: Spoofing, Tampering, Repudiation, Information disclosure, Denial of service and Elevation of privilege.
- Learn it in: Threat model an AI system
- Related: Threat modelling, MITRE ATLAS
- Structured output
- Making a model reply in an exact machine-readable shape, usually JSON that matches a schema, so code can use the answer directly.
- Learn it in: Reading and shaping JSON, Extracting data from documents, Triage for email and tickets, Get structured JSON output, Guardrails
- Related: JSON Schema, Tool calling
- Sub-agent
- An agent started by another agent to handle one part of a task in its own context, returning only a summary.
- Learn it in: Agent SDKs and sub-agents, Context engineering, Multi-agent systems, Advanced harness engineering, Harnesses for long-running agents
- Related: Agent, Context engineering, Multi-agent system
- Supervised fine-tuning (SFT) (also: SFT, instruction tuning)
- Training a model on examples of a prompt and the ideal answer, so it learns to follow instructions and reply like an assistant. Also called instruction tuning.
- Learn it in: How a chat model is made: pre-training, SFT and preference tuning, Building a training dataset: quality, coverage and quantity
- Related: Fine-tuning, Base model, RLHF (reinforcement learning from human feedback)
- Supervised learning
- Learning from examples that come with correct answers (labels), to predict the answer for new examples.
- Related: Label, Unsupervised learning
- Synthetic data
- Training or test examples made by a model or a program instead of collected from real people. It is cheap and can fill gaps, but it must be checked and mixed with real data.
- Learn it in: Building a training dataset: quality, coverage and quantity, Synthetic data and distillation: training with model-made data
- Related: Distillation, Model collapse, Training
- System prompt
- Standing instructions sent before the conversation that set the model's role, rules and output format.
- Learn it in: System prompts and prompt design
- Related: Prompt, Prompt injection
T
- Task vector (also: delta parameters)
- The change fine-tuning made to a model: the fine-tuned weights minus the base weights. Adding or removing task vectors adds or removes skills.
- Learn it in: Model merging: combining fine-tuned models into one
- Related: Model merging, Fine-tuning
- TCP/IP (also: TCP, UDP, IP address)
- The family of protocols the internet runs on. IP addresses and routes packets between machines; TCP adds reliable, ordered delivery; UDP sends quickly without delivery guarantees.
- Learn it in: Networking for security
- Related: Port, DNS
- Temperature
- A setting that controls how random a model's word choices are. Low gives steady, repeatable answers; high gives more varied ones.
- Learn it in: How a model picks its next word: temperature, top-p and top-k
- Related: Sampling
- Tensor
- A multi-dimensional array of numbers, the basic data structure in deep learning libraries such as PyTorch.
- Learn it in: NumPy and plotting, PyTorch
- Related: Vector, Matrix
- Tensor parallelism (also: intra-operator parallelism)
- Splitting each layer's matrix work across several GPUs that work at the same time, so a model too big for one GPU can run, often faster.
- Learn it in: Running models on many GPUs: replicas, tensor and pipeline parallelism
- Related: Pipeline parallelism, Data parallelism
- Terminal (also: command line, shell)
- A text window where you type commands to run programs and manage files.
- Learn it in: Set up Python and run your first file, The terminal and Git
- Related: Git
- Test fixture
- Setup that tests share, such as sample data or a temporary database, created before a test and cleaned up after. In pytest, a function marked @pytest.fixture.
- Learn it in: Writing tests with pytest
- Related: pytest, Unit test
- Test-time compute (also: inference-time compute)
- Spending more computing at answer time to get a better answer: letting a model reason for longer, or generating several answers and picking the best one.
- Learn it in: Reasoning models and test-time compute: when to let the model think longer
- Related: Reasoning model, Step-by-step reasoning, Best-of-N
- Text-to-speech
- Turning written text into natural-sounding spoken audio. Voice agents use it to say their replies.
- Learn it in: Build AI voice agents with Vapi: phone and web assistants that talk
- Related: Voice agent, Speech-to-text
- Text-to-video
- Generating a video clip from a written description, and often from reference images too. Google's Veo is an example.
- Learn it in: Create cinematic AI videos with Google Veo and Flow
- Related: Diffusion model
- Threat actor (also: Attacker)
- A person or group that may attack a system, such as a criminal gang, an insider, a hacktivist or a state-backed group, each with its own motives and skills.
- Learn it in: How cyber attacks work
- Related: Attack surface, CIA triad
- Threat modelling
- A structured way to find what can go wrong in a system before attackers do: draw how data flows, list threats to each part, rank them and decide the controls and tests for each.
- Learn it in: Threat model an AI system
- Related: STRIDE, MITRE ATLAS, Risk log
- Throughput (also: requests per second, RPS, QPS)
- How much work a system finishes in a period of time, such as requests per second. It is different from latency, which is how long one request takes.
- A toll plaza with ten booths has more throughput than one with two, even if each car takes the same time at its booth.
- Learn it in: Scaling and load balancing, How to approach system design, Message queues and background work, Inside LLM inference: prefill, decode and the KV cache
- Related: Latency, Horizontal scaling
- Time to first token (TTFT) (also: TTFT)
- How long a user waits before the first word of a reply appears. It depends mainly on prompt length and how busy the server is.
- Learn it in: Inside LLM inference: prefill, decode and the KV cache
- Related: Prefill, Latency
- TLS (also: Transport Layer Security, SSL, HTTPS)
- The protocol that encrypts traffic between a client and a server, such as a browser and a bank website, and uses certificates to prove the server's identity. HTTPS is HTTP over TLS.
- Learn it in: Networking for security, Cryptography basics
- Related: Encryption, Hashing
- Token
- The unit of text a model reads and writes: a whole word, part of a word or a punctuation mark. Prices and limits are counted in tokens.
- A rough rule for English is that 100 tokens are about 75 words. Hindi and Tamil text usually needs more tokens for the same meaning, so it costs more.
- Learn it in: Call an LLM API from code, Tokens and the context window, AI in Indian languages: tokens, quality gaps and what to use, Transformers, Caching and speed
- Related: Context window, Tokenizer
- Token fertility (also: tokens per word)
- How many tokens a tokenizer needs, on average, for one word of a language. Hindi or Tamil often need several times more tokens than English for the same meaning, which raises cost and delay.
- Learn it in: AI in Indian languages: tokens, quality gaps and what to use
- Related: Tokenizer, Token
- Tokenizer
- The part of a model's toolkit that splits text into tokens and turns tokens back into text.
- Learn it in: Tokens and the context window, AI in Indian languages: tokens, quality gaps and what to use, Perplexity and cross-entropy: how language models are measured
- Related: Token
- Tool calling (also: function calling)
- Letting a model ask your code to run a function, such as a search or a database lookup, by replying with the function's name and arguments.
- Learn it in: How AI products work, Build AI voice agents with Vapi: phone and web assistants that talk, The agent loop, Designing tools for agents, MCP servers and clients, Tool calling
- Related: Agent, JSON Schema, Model Context Protocol
- Tool poisoning (also: MCP tool poisoning, tool description poisoning)
- Hiding malicious instructions or behaviour in an AI agent's tool: in its description, code or results. Because the model trusts tool descriptions, hidden text there can steer the agent to misuse its tools.
- Learn it in: Securing agent tools and MCP servers
- Related: Model Context Protocol, Prompt injection, Confused deputy, Excessive agency
- Tool schema (also: tool definition)
- The name, description and parameter definitions of a tool that a model reads to decide when and how to call it.
- Learn it in: Designing tools for agents, Tool calling
- Related: Tool calling, JSON Schema
- Top-k sampling
- A sampling setting that keeps only the k highest-scoring next-word options and picks from them. It always keeps the same number of options, however sure or unsure the model is.
- Learn it in: How a model picks its next word: temperature, top-p and top-k
- Related: Sampling, Temperature, Top-p (nucleus) sampling
- Top-p (nucleus) sampling
- A sampling setting that keeps only the smallest group of next-word options whose chances add up to p, such as 0.9, and picks from those. The group gets bigger when the model is unsure and smaller when it is sure.
- Learn it in: How a model picks its next word: temperature, top-p and top-k
- Related: Sampling, Temperature, Top-k sampling
- Total cost of ownership (also: TCO)
- Everything a system costs over its life: model or GPU costs, infrastructure, engineering, support, compliance and the cost of changing it later.
- Learn it in: Make, Zapier and Power Automate, Feasibility and build versus buy, Measuring return on investment, Pricing AI features, Graph RAG, Cost of ownership, Talking to executives
- Related: Cost per request
- Tracing (also: traces)
- Recording every step of one request, including prompts, tool calls, retrieved passages, timings and costs, so you can see what happened.
- Learn it in: Audit logs, Tracing LLM apps, Logs and traces for agents, Agent SDKs and sub-agents, Multi-agent systems
- Related: Observability, Evals
- Training
- The process of adjusting a model's weights using examples until its outputs are good enough.
- Learn it in: How fine-tuning works: data, training and what can go wrong
- Related: Model, Weights, Fine-tuning
- Training-data extraction (also: memorisation, data regurgitation)
- Getting a model to repeat parts of its training data word for word, such as names, phone numbers or secret keys it memorised.
- Learn it in: Model extraction and privacy attacks
- Related: Membership inference, Model extraction, Personal data
- Trajectory
- The full sequence of an agent's steps on a task: its reasoning, tool calls, results and final output. Agent evals often score the trajectory, not just the answer.
- Learn it in: Evaluating agents
- Related: Evals, Agent
- Transfer learning
- Starting from a model trained on a large general dataset and adapting it to your task with much less data.
- Learn it in: CNNs for images
- Related: Fine-tuning, Convolutional neural network
- Transformer
- The neural network architecture, built on attention, behind modern language models and many vision and speech models.
- Learn it in: Transformers
- Related: Attention, Large language model
- Trigger
- The event that starts an automation, such as a new email, a form submission, a schedule or a webhook.
- Learn it in: What automation is, Your first n8n workflow, Webhooks
- Related: Automation, Webhook
- Two pointers (also: two-pointer technique, sliding window)
- A technique that walks two positions through an array, often from both ends or at different speeds, to solve a problem in one pass instead of nested loops.
- To check if any two prices in a sorted list add up to ₹1,000, start one pointer at each end and move them inward: one pass instead of checking every pair.
- Learn it in: Sorting, binary search and two pointers
- Related: Binary search, Array
- Type hint (also: type annotation)
- A note in Python code saying what type a value should be, such as name: str. Python does not enforce it, but editors and checkers like mypy use it to catch mistakes.
- Learn it in: Python for production
- Related: Pydantic
U
- Unit test
- A small automated test that checks one function or piece of logic in isolation, so a later change cannot silently break it.
- Learn it in: Clean code and refactoring, Debugging systematically, Project tooling: environments, tests and logs, Writing tests with pytest, Code review: giving and getting good reviews
- Related: Regression test, pytest
- Unsupervised learning
- Finding structure in data without labels, such as grouping similar customers or reducing many features to a few.
- Learn it in: Unsupervised learning
- Related: Clustering, Supervised learning
- Uvicorn
- A fast server that runs Python async web apps such as FastAPI. Production set-ups run several Uvicorn workers, sometimes managed by Gunicorn.
- Learn it in: Serving an AI app
- Related: async and await, Endpoint
V
- Variable
- A name that refers to a value in a program, such as price = 250, so the value can be used and changed later.
- Learn it in: Variables, loops and functions in Python, Write your own functions
- Related: Python
- Vector
- An ordered list of numbers. In ML, each example (a customer, an image, a sentence) is usually represented as a vector of features.
- Learn it in: Maths for machine learning, NumPy and plotting
- Related: Matrix, Embedding
- Vector database (also: vector store)
- A database built to store embeddings and quickly find the ones closest to a query embedding. Examples include pgvector, Qdrant and Pinecone.
- Learn it in: Embeddings, Vector search
- Related: Embedding, Semantic search
- Vectorless retrieval (also: reasoning-based retrieval, tree search retrieval)
- Finding the right part of a document without embeddings, for example by having a model read the table of contents and choose which sections to open.
- Learn it in: Long documents without vectors
- Related: Retrieval, Retrieval-augmented generation
- Verifier
- A program or model that checks or scores an AI answer, such as running generated code against tests or a reward model rating replies, so the best answer can be kept.
- Learn it in: Catching hallucinations: why models make things up and how to detect it, Reasoning models and test-time compute: when to let the model think longer
- Related: Best-of-N, Reward model
- Vibe coding
- Building software by describing what you want to an AI coding assistant in plain words and reviewing what it writes, rather than typing most of the code yourself.
- Learn it in: Vibe coding with Claude Code: build real apps by describing them
- Related: Prompt engineering, Agent
- Virtual environment (also: venv)
- A separate, project-specific set of installed Python packages, so projects do not interfere with each other.
- Learn it in: Packages and virtual environments
- Related: pip
- Vision-language model (also: VLM, multimodal model)
- A model that takes images and text as input, so it can read screenshots, scanned pages, charts and photos.
- Learn it in: Images, voice and other AI tools, Multimodal RAG
- Related: Multimodal RAG
- vLLM
- An open-source engine for serving language models on GPUs fast and efficiently, with an OpenAI-compatible API.
- Learn it in: GPUs and quantization, Running models on many GPUs: replicas, tensor and pipeline parallelism, Open-weight models with vLLM
- Related: Open-weight model, Self-hosting
- Voice agent
- An AI that talks with people by voice, on the phone or the web. It turns speech into text, decides what to say with a language model, and speaks the reply.
- Learn it in: Build AI voice agents with Vapi: phone and web assistants that talk
- Related: Speech-to-text, Text-to-speech
- Voice cloning
- Using AI to copy a person's voice from short recordings, so a fake call sounds like a boss or a relative. Defend with call-backs on a known number and agreed code words, not by trusting how a voice sounds.
- Learn it in: AI-powered attacks and how defenders respond
- Related: Deepfake, AI-generated phishing, Text-to-speech
- Vulnerability
- A weakness in software, configuration or process that an attacker could use to break confidentiality, integrity or availability.
- Learn it in: How cyber attacks work, Web app security and the OWASP Top 10
- Related: Exploit, CVE
W
- Webhook
- A message one app sends to a URL you give it when something happens, such as "new order placed", so another app can react at once.
- Learn it in: What automation is, Automation with n8n: workflows and AI agents without heavy coding, Reading and shaping JSON, Webhooks
- Related: API, Automation
- Weights (also: parameters)
- The numbers inside a model that training adjusts. Together they hold everything the model has learned.
- Learn it in: LoRA, QLoRA and the GPU memory maths
- Related: Model, Training, Open-weight model
- Window function
- An SQL function that computes over a set of rows related to the current row, such as a rank or running total, without collapsing rows like GROUP BY does.
- Learn it in: SQL for engineers
- Related: Join
- Workflow
- A fixed sequence of steps your code controls, which may include model calls. More predictable than an agent.
- Learn it in: The agent loop
- Related: Agent, Automation
- Workflow automation
- Connecting apps so a trigger, such as a new form entry, runs a chain of steps on its own. Tools like n8n and Zapier do this, often with AI steps inside.
- Learn it in: Automation with n8n: workflows and AI agents without heavy coding
- Related: Automation, Webhook