Verifiable citations
Claude’s native citations return the exact quoted passage behind each claim. Click a citation and a side panel opens with that passage highlighted in its source document.
Answers only from your documents, and shows you the exact passage it used.
A website chatbot that answers from a company’s own documents, cites the exact quoted passage behind every claim, says plainly when it doesn’t know, captures leads and hands over to a human. One script tag installs it on any site.
Businesses want a chatbot that answers customers at 2 am, but generic bots make things up, can’t prove where an answer came from, and quietly go out of date when prices or policies change.
Claude’s native citations return the exact quoted passage behind each claim. Click a citation and a side panel opens with that passage highlighted in its source document.
Vector search (pgvector) and keyword search fused with reciprocal rank fusion, so exact terms like postcodes, fault codes and model numbers are found as reliably as meaning.
No relevant source, no answer. Unanswered questions are logged and grouped by topic in a knowledge-gap report, so the business knows which page to write next.
When a visitor wants a quote, the bot collects their details with a consent line. “Talk to a person” sends the full transcript to email and Slack.
Upload Markdown or PDF, or import a web page or sitemap. Re-indexing replaces a document atomically, so outdated content can never be retrieved.
Secure login, documents, conversations with their citations, leads, low-rated answers, the knowledge-gap report and brand settings, all in one place.
A 50-question tuning set and a 20-question held-out set (out-of-scope questions, multi-turn follow-ups and prompt injections included), graded for retrieval, citation accuracy, honest refusals and cost. The report will be published here with the first-run numbers.
Document text and visitor messages are data, never instructions. Uncited answers are treated as “I don’t know”.
The knowledge base contains planted instructions (a fake 50% discount, fake bank details) to prove the bot ignores them.
Cloudflare Turnstile, per-IP rate limits, message and turn caps, and a daily spend cap that switches to “please call us”.
The app role can only touch its own schema, over verified TLS. Admin credentials never reach the app.
The widget runs in an iframe, so it can’t clash with or break the host site, and works on mobile and with a keyboard.
Every answer records its tokens, cost and timing, with prompt caching on the fixed system prompt.
Tell me what it is. I’ll come back with whether AI is the right fix, what it would take, and what it would cost to run.
Replies within 24 hours · Free 30-minute discovery call
That’s enough for me to come back with an approach, a rough timeline and the questions that matter.