Terms, defined plainly.
The vocabulary that appears across our writing, defined without marketing gloss. Where a term is commonly misused, the definition says so.
Answer engine optimisation
Writing and structuring content so that AI answer systems can find, quote and attribute it. In practice it means leading with a direct answer, being specific enough to quote, and sourcing factual claims.
Chunking
Splitting content into retrievable passages. Where the boundaries fall matters more than how large the pieces are, because a boundary that cuts an answer in half produces two passages, neither of which answers the question.
Context window
The amount of text a language model can consider at once, including the supplied passages, the conversation history and the answer. Larger is not automatically better, because irrelevant context dilutes the relevant part.
Credit
A unit of assistant activity used for billing, typically corresponding to answering a question. Its exact mapping is a product decision and should be stated plainly on a pricing page rather than left to be discovered.
Deflection rate
The share of conversations resolved without a human. Widely used as a headline metric and misleading on a marketing site, where reaching a person is sometimes the best available outcome.
Embedding
A representation of text as a list of numbers, arranged so that similar meanings produce similar numbers. Captures topic and paraphrase well. Handles negation, quantity and recency poorly.
Grounding
Constraining a model to answer from supplied source material rather than from general knowledge. Grounding is what makes an assistant answer about your product instead of about products in general.
Hallucination
A confident, fluent answer that is not supported by any source. On a website assistant the usual cause is retrieval failure rather than model failure: the model was handed the wrong passage and used it as instructed.
Handoff
Transferring a conversation to a person. A good handoff carries the full transcript, the pages viewed, what the assistant attempted and why it escalated. Without that context, the visitor starts over.
Hybrid retrieval
Combining vector similarity search with keyword search. Covers cases neither handles alone, particularly questions whose answers hinge on a specific number or exact term.
Indexing
Processing content into a searchable store. Includes fetching, extracting text, chunking, embedding and storing. Content that has changed since indexing produces stale answers until reindexed.
Intent
What a visitor is trying to establish. Phrasing predicts it more reliably than page location: capability questions, procedure questions, troubleshooting and risk assessment each have recognisable shapes.
Knowledge boundary
The edge of what an assistant has sources for. A defined boundary produces honest refusals. An undefined one produces confident answers about things nobody wrote down.
Latency
Time between a visitor sending a message and receiving an answer. Model time usually dominates. Streaming reduces perceived latency without reducing total latency, and perception is what visitors experience.
llms.txt
A plain text file at a site's root describing what the site is and where its primary content lives, intended for large language models. A different thing from a marketing page about AI.
Multi model routing
Choosing which model handles a given query based on cost, latency and task difficulty, rather than sending everything to one. Also provides a path for provider failover.
Rate limiting
Capping requests per session, per origin and globally. A public endpoint that triggers paid API calls has costs controlled by whoever visits, which makes the global limit the only one that guarantees a bounded bill.
Reranking
A second pass that reorders retrieved candidates using a more expensive and more accurate method. Useful when first pass retrieval returns roughly the right area but the wrong specific passage.
Retrieval augmented generation
Looking up relevant passages at question time and supplying them to a model along with the question. Converts the task from recall to reading comprehension, which is why it works on content the model has never seen.
Refusal
Declining to answer because no source covers the question. A designed feature rather than a failure. On a public site, a wrong answer about pricing or policy costs more than an honest refusal.
Source attribution
Showing which indexed content an answer came from. Lets a visitor verify and lets you diagnose whether a bad answer was a retrieval problem or a generation problem.
Structured data
Machine readable markup describing a page's content, usually JSON-LD. Should represent what is actually visible on the page. Marking up content that is not present is the kind of shortcut that ages badly.
Unanswered rate
The share of conversations where the assistant could not answer. The direct measure of content gap size. A very low rate is not good news; it usually means the refusal threshold is too permissive.
Vector search
Finding stored passages whose embeddings are closest to a question's embedding. Fast and fuzzy. A filter that narrows thousands of candidates to a handful, not a decision about which one is correct.