Product · AI answers
Answers from your content. With the source attached.
Retrieval-augmented answers over the same indexes your search uses. Every answer links to the documents it came from; retrieved text is screened for prompt injection, and answers are checked for grounding in the sources.
How an AI answer is built
Every answer goes through the same six steps. The model only sees what search retrieved — and the visitor sees where it came from.
- 01
Question
“Does this jacket run small?” — in the visitor's own words.
- 02
Rewrite
Turned into a search query, using earlier messages to resolve “this”.
- 03
Retrieve
Hybrid search over your index: product page, size guide, reviews.
- 04
Filter
Access rules applied; instruction-like text in documents is screened out.
- 05
Answer
Written only from the retrieved passages. No passages — no invented answer.
- 06
Cite
Every source linked; clicks on citations tracked like search clicks.
How it answers
Retrieval first
The question is rewritten into a search, run against your index with the same tenant and access filters, and only the retrieved passages are given to the model.
Citations
Each answer returns its sources with titles and links; click events on citations are tracked like search clicks.
Conversation memory
Follow-up questions are resolved against the thread, stored per index.
Streaming
Server-sent events for token-by-token answers in the widget or your own UI.
Safety and control
Permission-aware
Documents with ACL are retrieved only for users allowed to read them, so an answer cannot quote a hidden document.
Injection screening
Retrieved text is scanned for instruction-like patterns in several languages before it reaches the model.
Feedback and evaluation
Thumbs up/down per answer, plus golden question sets scored after every content change.
Coming next: GraphRAG
Questions across many documents
Will answer broad questions by summarising communities of related entities and documents, not only the top passages.
Specifications
- Question length
- up to 1,000 characters
- Answer delivery
- single response or streamed (SSE)
- Conversation memory
- per thread, stored per index
- Document chunking
- 200–8,000 characters per chunk, configurable overlap
- Access control
- document ACL applied before retrieval
- Free plan
- 10,000 AI tokens / month
Before you start
- Content to answer from: help-centre articles, docs, product pages
- An LLM provider configured on the platform (answers fall back to extracts without one)
- 10–20 real customer questions to check answers before launch
How to start
- 1
Index your content
Crawl the help centre or docs, upload Markdown or JSON, or reuse the product index.
- 2
Ask
One call returns the answer, citations and a conversation id for follow-ups.
await client.assistantAsk("help-center", { question: "How do I return an order?" });