Knowledge
Train your agent on your website, docs, help centre, and FAQs so every reply is grounded in your business rather than guessed.
Your agent does not answer from general world knowledge. It answers from your knowledge. This page explains the implemented website/FAQ path and the broader source-management design.
What you can train on today
Website URLs and hand-written FAQs, with chunking, embedding, and hybrid retrieval. File upload, progressive onboarding crawls, viewable answer citations, vertical templates, and interview-driven setup are being built.
What you can connect
| Source type | What it is | Notes |
|---|---|---|
| Website | A crawl of your public site, docs, or help centre | Crawled and cleaned to plain text; navigation, footers, and boilerplate are stripped |
| Files (future) | Uploaded documents (product sheets, policies, price lists, catalogs) | The Phase 2 API does not accept file uploads |
| FAQs | Hand-written question-and-answer pairs | Each Q&A is kept whole as one unit, never split |
| Learned answers | Answers your team gives during a handoff, promoted to FAQs | Roadmap — needs an approval step before anything is trusted |
Knowledge is owned by your organisation, and each agent is linked to the sources it is allowed to read. Two agents in the same org can have different knowledge — a support agent reading the help centre, a sales agent reading pricing and case studies — without duplicating anything.
How ingestion works
Fetch
For a website, a crawler pulls the pages and converts them to clean markdown — main content only, so menus and cookie banners never become "knowledge". Live Firecrawl execution remains a provider-acceptance check. Direct file ingestion is future work.
Chunk
Each document is split into passages of roughly 400–500 tokens with a small overlap, cut on heading boundaries and carrying their heading breadcrumb so a passage still knows which section it came from. FAQs are atomic: one chunk per question, never split mid-answer.
Embed and index
Every chunk gets a vector embedding and a full-text index entry. The language of each document is detected at ingest so full-text search works on non-English content instead of quietly failing.
Ready
Chunks become retrievable for the agents linked to that source. Progressive homepage-first onboarding while the rest of a crawl continues is a future enhancement, not Phase 2 behavior.
How a reply gets grounded
On every incoming message, before the model writes a single word, CoPhrase assembles the turn's context:
- your persona and instructions (see persona)
- your structured business facts
- the top few knowledge chunks retrieved for this specific question
- the contact's card and history
- guardrail directives
Retrieval is hybrid. Two searches run against your knowledge — a vector search (semantic: "how much is it" finds a page titled Pricing) and a full-text search (lexical: exact product names, SKUs, and error codes that embeddings blur). The two ranked lists are fused, and the best five to seven passages are injected into the turn.
Ties are broken on per-source authority and freshness that you control, not on a hard-coded ordering. A curated FAQ usually beats a crawled page — but not if the FAQ is two years old and the page is not.
The promise, and the limit
Phase 2 injects retrieved chunks into the model context, but it does not yet persist the cited chunk IDs on each reply or expose per-chunk citation counts. Reply-level source traceability and citation analytics are future product surfaces.
The agent's job is to answer from your data, not to guess. Retrieval confidence is banded:
| Confidence | Behavior |
|---|---|
| Strong | Answers from the retrieved passages |
| Weak | Answers cautiously, and is more likely to escalate |
| None | Does not free-generate. It sends a templated "I can't answer that" and escalates to a human |
Prices, policies, and opening hours are held to a stricter rule: they must come from your structured business profile or from a cited chunk. The agent is not permitted to invent them.
Grounding is not the same as being right
Grounding guarantees the agent answers from your source. It does not make your source correct. If your pricing page is stale, the agent will confidently quote stale prices — and it will be traceable back to the page that told it to. Wrong knowledge in, wrong answer out. This is why sync matters.
Keeping knowledge in sync
- Re-crawl a website source to pick up changes. Each document is content-hashed, so a re-sync only re-processes pages that actually changed — the rest are skipped, which keeps re-syncs fast and cheap (see billing).
- Edit FAQs directly for anything that changes faster than your website does, or that was never on your website in the first place.
- Future: watch what is being cited. Persisted reply citations and chunk-level counts will show which passages are carrying conversations once that product surface lands.
- Watch what the agent could not answer. Escalations and "none"-confidence conversations are your knowledge gaps, surfaced as work rather than lost.
No website yet?
A crawl or hand-written FAQ is the Phase 2 path. Vertical templates, a guided business interview, and uploaded files are future onboarding options.
Multilingual
Language is detected per document at ingest, and indexed accordingly. The agent mirrors the language of the incoming message — a Spanish question against an English knowledge base still gets a Spanish answer grounded in the English source.