Glossary

Shared vocabulary, defined once.

Terms used consistently across every guide, workflow pattern, and research note on this site.

AI agent
A model-driven program that plans steps, calls tools, and works toward a goal instead of answering a single prompt.
Agents differ from chat completions mainly in control flow: they decide what to do next. Web access is one of the most common tools given to an agent, and also one of the most common sources of unreliable behaviour.
Related: agentic-workflow / web-retrieval / citation
Web retrieval
Fetching the content of a known URL and returning it in a form a program or model can use.
Retrieval covers the request, any rendering required, and the decision about what part of the response is actually content. It is the narrowest of the web-data operations and the easiest to reason about.
Related: content-normalization / javascript-rendering
Web crawling
Following links or a sitemap to retrieve many related pages under a defined scope.
Crawling adds scope rules, deduplication, and politeness constraints on top of retrieval. Most crawl problems are scope problems: too broad and you drown, too narrow and you miss the pages that mattered.
Related: sitemap / canonical-url / knowledge-base
Web scraping
Pulling specific values out of pages, usually into fields defined in advance.
Classic scraping binds to markup structure, which breaks when layouts change. Schema-driven extraction binds to meaning instead, which is more tolerant but introduces its own ambiguity.
Related: structured-extraction / schema
Browser automation
Driving a real browser to click, type, scroll, or authenticate in order to reach content.
Use it when content only exists after interaction. It is slower, more fragile, and more expensive than retrieval, so treat it as an escalation rather than a default.
Related: javascript-rendering / web-retrieval
Structured extraction
Converting page content into records that match a defined schema, with types and optional fields.
Good extraction records what it could not find rather than guessing. A null with a reason is more useful downstream than a plausible invention.
Related: schema / source-provenance
Schema
A declaration of the fields, types, and requirements an extracted record must satisfy.
Schemas are a contract between messy input and dependable output. They make validation possible and turn 'the page changed' into a visible failure instead of a silent one.
Related: structured-extraction / citation
Markdown
A lightweight text format that keeps headings, lists, links, and tables without layout markup.
For most language-model tasks, Markdown is a better carrier than HTML: it preserves the structure that aids comprehension and discards the structure that only serves rendering.
Related: content-normalization / chunking
RAG
Retrieval-augmented generation: supplying a model with retrieved passages so it answers from supplied context.
RAG quality is dominated by what goes into the index. Retrieval strategy matters, but unusable source content limits the ceiling before any ranking work begins.
Related: embedding / chunking / knowledge-base
Embedding
A numeric vector representing text, used to find passages that are semantically similar to a query.
Embeddings find related language, not correct answers. They are one retrieval signal among several, and they degrade when chunks mix unrelated topics.
Related: rag / chunking
Chunking
Splitting a document into retrievable pieces sized for indexing and for a model's context window.
Chunk on document structure where possible — headings, sections, steps — rather than fixed character counts. Carry the page title and section path into each chunk.
Related: rag / markdown
Source provenance
The retrieval record attached to a piece of content: URL, retrieval time, and how it was obtained.
Provenance is what makes an AI output auditable. Without it, a wrong answer cannot be traced to a wrong source, only to the model.
Related: citation / structured-extraction
Citation
A link from a specific claim to the specific source passage that supports it.
A citation attached to a paragraph is weaker than one attached to a sentence. Quality is judged by whether a reader can verify the claim quickly, not by the presence of a link.
Related: source-provenance / ai-agent
Semantic change detection
Deciding whether a page difference is meaningful rather than cosmetic.
Byte diffs alarm on rotating banners and session tokens. Semantic detection compares normalized content or defined fields and ignores the rest.
Related: content-normalization / web-retrieval
Sitemap
A published list of a site's URLs, usually at /sitemap.xml, intended for crawlers.
Sitemaps are the cheapest possible discovery mechanism for a documentation crawl, but they are frequently stale or partial. Treat them as a starting set, not as truth.
Related: web-crawling / canonical-url
Canonical URL
The address a site declares as the preferred version of a page that is reachable at several URLs.
Respecting canonicals is the single highest-value deduplication rule in a crawl. Ignoring them is how a 400-page documentation set becomes 4,000 near-identical chunks.
Related: web-crawling / sitemap
JavaScript rendering
Executing a page's scripts so that client-generated content exists in the retrieved document.
Rendering costs time and compute. Check whether the data arrives in an underlying JSON response or an embedded state blob before paying for a full render.
Related: browser-automation / web-retrieval
Content normalization
Stripping navigation, chrome, and boilerplate so the retained text is the page's actual content.
Normalization is where most quality is won or lost. Over-aggressive cleaning removes tables and code; under-aggressive cleaning floods the index with cookie banners.
Related: markdown / rag
Agentic workflow
A multi-step process where a model chooses actions, calls tools, and reacts to intermediate results.
The reliable ones constrain the shape of each step. Free-form autonomy over the open web tends to produce interesting demos and unpredictable products.
Related: ai-agent / citation
Knowledge base
A curated, maintained collection of content prepared for retrieval by people or systems.
A knowledge base is a maintenance commitment, not a one-time build. Recrawl cadence, removal of dead pages, and review dates matter as much as the initial ingest.
Related: rag / web-crawling