Helpin
Dark mode
Knowledge-base answers

Knowledge-base answers

The AI support agent answers from your content using retrieval-augmented generation (RAG): it retrieves relevant passages from your connected knowledge sources, then writes an answer grounded in them. It does not answer from general knowledge.

Which sources are used

All sources connected for support in Settings → Knowledge:

  • Docs spaces — Helpin documents (internal or external-capable spaces).

  • Help center articles — published public docs.

  • Website sources — public pages Helpin crawls from URLs you add.

Each source has a Company/Product context description that gives the AI background about what your product is and does — this context accompanies retrieval and improves answer framing.

Knowledge settings page with company context and content sources

How an answer is produced

  1. The customer's message is embedded and searched against your indexed content to find the most relevant articles and passages.

  2. Top passages are retrieved with their metadata (title, source, URL).

  3. The AI composes an answer grounded in those passages, in the conversation's context (previous turns, clarifications).

  4. Citations attach to the sources actually used.

  5. Confidence is evaluated — if the retrieval and answer confidence fall below your configured threshold, the AI escalates instead of answering.

Why answers stay grounded

  • The AI is instructed to answer only from retrieved content — not from its general knowledge.

  • When retrieval finds nothing relevant, the correct outcome is a handoff, not an improvised answer.

  • The answer text must be consistent with the cited passages; agents can inspect citations to verify.

Keeping content fresh

Retrieval quality is content quality:

  • Publish and update articles in Docs; changes propagate to the AI's index as content is republished.

  • Prefer a smaller number of clear, current articles over many overlapping drafts — overlapping or outdated content dilutes retrieval.

  • Use the Coverage Gaps view (see Coverage analysis) to find questions the AI couldn't answer — each gap is a content opportunity.

What RAG does not do

  • It does not answer account-specific questions ("what's my invoice status?") unless that data is in your content — those need human support.

  • It does not train on your conversations; answers come from your documents at question time.

  • It does not mix in public web content — only what you connected.

Testing your setup

  1. Ask a question in a test conversation that your docs clearly answer.

  2. Verify the reply cites the right article.

  3. Ask a question your docs don't cover — verify it escalates rather than invents.

Was this article helpful?