A RAG chatbot for websites, docs, and help centers that retrieves from your own content instead of guessing. ChattyBox answers questions like "What is included in the Pro plan?" from your published pages and cites the source — with no custom retrieval infrastructure to build or host.
RAG combines a language model with retrieved source content. Before generating a response, the chatbot searches your published pages for passages that match the question.
If your priority is answer trust, pair this retrieval workflow with a grounded AI chatbot that avoids invented answers through source constraints and fallback behavior.
When a visitor asks a question, ChattyBox searches indexed website, docs, help, and CMS content for relevant passages.
Those source passages are added as context so the model has something specific to answer from.
The model generates a concise answer and can include source links so visitors can verify the details.
A dependable RAG system is a chain of testable stages, not a single model call. This reference architecture separates content preparation, retrieval, generation, and citation so teams can locate failures and re-index changed sources without retraining a model.
Reviewed when retrieval behavior changes and at least every six months.
| Stage | What happens | Control to verify |
|---|---|---|
Stage 1. Crawl | What happens Discover allowed pages from URLs and sitemaps. | Control to verify Expected URLs are present; excluded and private paths are absent. |
Stage 2. Parse | What happens Extract headings, prose, lists, tables, code, canonical URLs, and metadata. | Control to verify Navigation noise is removed without dropping commands or version labels. |
Stage 3. Chunk | What happens Split content into coherent passages that retain page and section identity. | Control to verify Each passage is understandable alone and preserves prerequisites. |
Stage 4. Embed | What happens Represent passages for semantic matching and store source metadata. | Control to verify Every vector maps to one current, permitted source passage. |
Stage 5. Retrieve | What happens Rank a small set of passages against the visitor question. | Control to verify Recall@k tests put the gold passage in the context set. |
Stage 6. Prompt | What happens Combine the question, passages, source metadata, and fallback rules. | Control to verify Instructions cannot expand permissions or override fallback rules. |
Stage 7. Answer | What happens Generate only supported details or decline when evidence is insufficient. | Control to verify Material claims follow the context; missing evidence triggers fallback. |
Stage 8. Cite | What happens Attach the canonical page and section supporting the response. | Control to verify Links resolve, match the version, and directly support the claim. |
A source change is not retrievable until the affected page is crawled, parsed, chunked, and embedded again. Track source revisions, replace stale chunks instead of appending duplicates, remove deleted pages, and test representative answers after re-indexing. Set recrawl frequency from the shortest acceptable stale-answer window for pricing, policy, and versioned docs.
Public pages can share one crawl scope. Private, tenant, employee, or account data needs authenticated ingestion and retrieval-time authorization before passages enter the prompt. Robots rules are not access control. Never index secrets, and test that unauthorized identities cannot retrieve titles, snippets, URLs, or answer details from restricted sources.
Use a versioned question set with known source passages, expected claims, and expected fallback cases. Run it after source, chunking, embedding, ranking, or prompt changes, then inspect failures by stage.
| Signal | How to measure it | Failure it exposes |
|---|---|---|
Signal Recall@k | How to measure it Share of answerable questions where a gold passage appears in the top k results. | Failure it exposes Missing crawl coverage, weak chunks, vocabulary mismatch, or stale embeddings. |
Signal Ranking quality | How to measure it Record reciprocal rank or nDCG for judged relevant passages. | Failure it exposes A relevant page exists but noise or an obsolete version outranks it. |
Signal Claim support | How to measure it Review every material claim against the passages supplied to the model. | Failure it exposes The answer combines assumptions with evidence or overstates the source. |
Signal Citation accuracy | How to measure it Verify that each citation resolves and directly supports its attached claim. | Failure it exposes The right answer links to a merely related page or the wrong version. |
Signal Fallback precision and recall | How to measure it Include unanswerable and conflicting-source questions; score correct declines and missed answers. | Failure it exposes The bot guesses without evidence or refuses despite adequate evidence. |
Signal Freshness lag | How to measure it Measure time from a published source change to a correct retrieved answer. | Failure it exposes Recrawl delay, duplicate stale chunks, or failed deletion propagation. |
| Method | Best fit | Important limit |
|---|---|---|
Method Keyword or site search | Best fit Users want ranked pages and will read the source themselves. | Important limit Does not synthesize an answer and can miss semantic matches. |
Method RAG | Best fit Knowledge changes, citations matter, and answers span a few passages. | Important limit Quality depends on crawl coverage, retrieval, permissions, and source quality. |
Method Fine-tuning | Best fit Teach stable style, format, classification, or task behavior. | Important limit Not a reliable replacement for current facts, citations, or source deletion. |
Method Long-context prompting | Best fit A bounded document set fits in context and every item is relevant. | Important limit Cost and noise grow with context; permissions and freshness still need controls. |
Suppose the indexed installation page says the widget script belongs immediately before the closing body tag. Retrieval should select that passage, the prompt should prohibit unsupported placement advice, and the answer should state the instruction with the installation page attached as evidence.
Continue with the guides to source citations, Docusaurus AI Chatbot for Source-Cited Docs, developer documentation chatbots, and the MkDocs AI Chatbot for Source-Cited Docs.
Your website, docs, help center, and CMS pages are indexed so the chatbot can search the content you already publish.
Embeddings help match visitor questions to relevant source passages even when the wording is different.
For each question, ChattyBox retrieves a small set of relevant passages instead of asking the model to answer from memory.
The answer step is constrained to the retrieved context and configured to avoid unsupported claims.
Every response includes links to source pages, so users can verify the information.
When relevant content is missing, the bot can fall back instead of improvising an unsupported answer.
A generic chatbot answers from broad model memory and can sound confident about things it does not know. A RAG chatbot retrieves from your pages first, so the answer is tied to content you published and can be checked.
When citations are the priority, pair retrieval with an AI chatbot with source citations that links every answer to its source page.
Connect your existing site content, test answer quality, and embed a source-cited chatbot when the responses look right.
Start Free - No Credit CardWe use optional analytics and tag-management tools to understand site use. Choose whether to allow PostHog and Google Tag Manager. Turning analytics off reloads this page so the change takes effect cleanly. Essential site functionality and error monitoring are not controlled by this choice. Read our privacy policy.
