The practical AI encyclopedia
Clear answers to the AI questions that shape real decisions. Definitions, how-to guides, and “when should I?” calls, with sources and a visible review date.
An LLM is a model trained to estimate and generate sequences of tokens. That lets it write, summarize, translate, classify, and reason over language, but it does not make every generated claim a verified fact.
Speech-to-text, or STT, converts spoken audio into written text. It can run on recorded files, live streams, or short clips, and may add punctuation, speaker labels, timestamps, and confidence estimates.
Text-to-speech, or TTS, converts written text into playable synthetic speech. Applications use it for voice interfaces, accessibility, narration, alerts, and audio versions of written content.
A vision-language model can process visual inputs such as images or documents together with language. It can describe, answer questions about, compare, or extract information from visual material.
STT turns speech into text, VLMs interpret visual input, an LLM reasons over the combined context, and TTS turns an approved response back into audio. Each component solves a different part of the interaction.
Use speech-to-text when audio is the source of useful information and converting it into searchable, editable text saves more work than manual transcription and review.
Use text-to-speech when people benefit from hearing approved text, or when a voice interface removes friction that a screen would create. Do not add synthetic narration just because audio is possible.
Use OCR when the job is reliable character extraction from clean text. Use a VLM when the task depends on layout, visual relationships, a chart, a screenshot, or a question about what the image means.
Build voice-to-content as a chain: transcribe the recording, clean and label the transcript, extract claims, draft assets, verify facts, and have a person approve the final text before publishing.
Build image-to-insight with an explicit question, a vision model or OCR step, extracted evidence, a human review, and a final decision that keeps observation separate from interpretation.
Agent memory is the information an AI agent can carry forward while working. It may be short-term task context, saved user preferences, past events, or searchable knowledge, and each kind needs different retention and privacy rules.
Use hybrid vector and graph search when a question needs both relevant passages and explicit relationships. Vector search finds the right language; a graph supplies ownership, dependencies, lineage, or multi-hop connections.
Build a knowledge graph by starting with the questions it must answer, defining a small vocabulary, extracting entities and relationships from trusted sources, validating them, and adding provenance before connecting an AI system.
Generate synthetic data safely by defining the use case, minimizing sensitive inputs, checking whether the output memorizes real records, comparing distributions and failure cases, and keeping a protected real evaluation set.
Protect private data by sending less of it, limiting who and which models can access it, isolating tenants, logging access, defining retention, and testing the system for accidental disclosure.
Deploy an AI system like a changing software system, not a finished prompt: version the prompts and models, test representative tasks, protect data, monitor cost and quality, and keep a rollback path.
Monitor an AI system across four layers: service health, model behavior, data and retrieval quality, and user outcomes. Logs alone are not monitoring unless they help you detect and repair failures.
Use a knowledge graph when the important part of an answer is how entities relate: ownership, dependencies, history, eligibility, or multi-step paths. Use simpler retrieval when the job is mostly finding relevant passages.
Use synthetic data when you need safe, repeatable examples for development, testing, or rare cases and can measure whether the generated data resembles the behavior that matters.
Use a local model when keeping data inside your controlled environment is more important than hosted-model convenience and you can accept the operational cost of running and evaluating it.
Use hybrid search when queries mix exact identifiers or terms with natural-language descriptions. It is especially useful for support, documentation, catalogs, and enterprise search with filters.
Add AI observability before an AI feature handles real users or important data. The more models, tools, retrieval steps, and side effects you introduce, the earlier you need a trace of what happened.
A knowledge graph represents entities and the relationships between them so an AI system can reason over connected facts instead of treating every document as an isolated passage.
Synthetic data is artificially generated data designed to resemble real data for testing, training, development, or privacy-sensitive work. It is useful only when its relationship to reality is measured.
AI data privacy is the set of controls that determine what information an AI system receives, who can access it, how long it is retained, whether it is used for training, and how it can be deleted.
Hybrid search combines lexical search, which matches words, with semantic search, which matches meaning. It is useful when a query includes both exact identifiers and natural-language intent.
AI observability is the ability to see and understand what an AI system received, which model and tools it used, what it returned, what it cost, and where the workflow failed.
Semantic search looks for meaning and intent rather than only matching the exact words in a query. It is useful when people describe the same thing in different language.
A context window is the amount of text and other input a model can consider during one request, including your instructions, conversation, retrieved documents, and the answer it generates.
A multimodal AI model can work with more than one kind of input or output, such as text and images, audio and text, or video and spoken instructions.
Create a reusable prompt by separating the task, variables, rules, source material, output format, and evaluation criteria. A template should reduce repeated thinking, not hide the decisions a user still needs to make.
Test an AI agent as a system, not just as a conversation. Check the plan, each tool call, the final result, the failure response, and what happens when the agent lacks permission or evidence.
Use a vector database by embedding the documents and the user’s query into the same space, retrieving nearby candidates, then applying permissions and relevance checks before showing results.
Reduce hallucinations by limiting what the model can claim, grounding answers in reviewed sources, testing missing-evidence cases, and measuring factual failures instead of relying on confident tone.
Use semantic search when people describe the same idea in many ways. Use keyword search for exact identifiers, names, and filters. Use hybrid search when both kinds of precision matter.
Use prompt engineering first when the task, rules, and output contract are still changing. Consider fine-tuning only after the behavior is stable, examples are strong, and prompting has reached a measured limit.
RAG is like giving an AI an open-book test: it searches a chosen set of documents for useful passages, then uses those passages to answer instead of relying only on memory from training.
A vector database stores numerical representations of information so an application can find items that are meaningfully similar to a query, even when they do not share the same words.
An embedding is a numerical representation of an item’s meaning or features. Applications compare embeddings to find related text, images, products, or examples.
Prompt engineering is the deliberate design and testing of the instructions, context, examples, output format, and constraints given to an AI model for a specific job.
Agentic AI describes systems that can plan or choose intermediate actions, use tools, inspect results, and continue toward a goal instead of producing one response and stopping.
Implement RAG as a small pipeline: prepare trustworthy documents, retrieve relevant passages, pass them to the model with a grounded instruction, and evaluate retrieval and answer quality separately.
Evaluate a model on a versioned set of real tasks with a defined approval bar. Score the output, correction effort, safety, latency, and total cost instead of treating a benchmark score as a production decision.
Improve retrieval by finding out which layer fails first. Fix source quality and chunking before adding a more expensive model or search service.
A good RAG prompt makes the evidence boundary explicit: what the model may use, how it should cite it, what it should do when evidence is missing, and what the final answer must look like.
Use a workflow when the steps are known and repeatable. Consider an agent when the path genuinely changes with the situation and the system can safely inspect results, choose a next action, and stop for approval.
Use a vector database when meaning-based similarity is central to the lookup. Use SQL or full-text search when exact filters, transactions, or precise terms matter. Many useful systems combine them.
An AI agent is a system that can decide which steps to take, use tools, inspect the results, and continue toward a goal. A chatbot answers a turn; an agent runs a bounded loop.
Build the smallest pipeline that can answer one defined class of questions from a controlled document set. Start with retrieval quality and citations before adding agents, reranking, or complex orchestration.
Fine-tune after you have a stable task, a strong example set, a repeatable evaluation, and evidence that prompting or retrieval cannot solve the failure. Fine-tuning is an optimization step, not a substitute for unclear requirements.
RAG is a way to let an AI answer from your own changing information without retraining the model. A support assistant, for example, can look up the current refund policy before answering instead of relying on whatever it learned during training.
Compare models on one repeated job, using the same inputs and the same approval bar. A model that looks cheaper per token can be more expensive if a person has to rewrite every result.
Start with RAG when the missing ingredient is current or private information. Consider fine-tuning when the model knows enough but repeatedly fails to behave the way your application requires. Most teams should test retrieval and better prompting before paying for a training pipeline.
A token is a small piece of text that an AI model reads or generates. Models use tokens to measure context, output limits, and often API cost; a token is not always a whole word.
AI inference is the process of running a trained model to produce an output for new input. Every chatbot answer, classification, image generation, and API prediction is an inference request.
AI guardrails are rules and checks that constrain an AI system before, during, or after a model call. They can block unsafe inputs, enforce an output shape, restrict tools, and stop unsupported results from being delivered.
A model router is a layer that decides which AI model should handle each request. It can send simple work to a cheaper model and reserve a stronger model for tasks that justify its cost.
A golden dataset is a curated set of representative inputs with trusted expected outcomes or grading rules. Teams use it to detect regressions when prompts, models, retrieval, or tools change.
Reduce token cost by removing repeated context, retrieving only what the task needs, caching stable instructions, routing simple jobs to cheaper models, and checking that the savings do not create more human correction.
Add guardrails around the risky boundaries of the workflow: validate inputs, limit context, constrain tools, check outputs, log decisions, and require approval for irreversible actions.
Detect hallucinations by comparing outputs with trusted evidence, checking claims individually, testing missing-information cases, and monitoring correction patterns in real use.
Choose a model router by the decisions it can make and prove: task classification, model selection, fallback behavior, cost controls, quality measurement, and logs that explain every route.
Use prompt caching when the same large instructions or context are sent repeatedly and the provider supports reliable cache behavior. It is valuable for cost and latency, not a replacement for good prompt design.
Use a model router when different requests genuinely need different models and the savings or reliability improvement can be measured. Avoid it when one model already handles the workload simply and well.
Use a smaller model when the task is narrow, the quality bar is measurable, and a larger model’s extra capability does not change the approved result. Test correction time, not just benchmark scores.
Add guardrails before real users depend on an AI feature, especially when it handles private data, produces public claims, calls tools, or can change money, accounts, content, or customer records.
Use keyword search when the exact term carries meaning, such as an order ID, product code, legal phrase, or date. Semantic search is not automatically better; it is better for a different kind of question.
Our standard
Each page is created for a real reader, grounded in primary sources, and reviewed for accuracy. We combine the direct answer with examples, trade-offs, and the next useful action.
See editorial standards ↗