Skip to main content

​
Retrieval keeps live answers current and accountable.

Callaro can retrieve knowledge during a live call so the AI answers with current, approved information rather than guessing from stale model memory. Agents attach retrieval in one of two ways:
  • Custom knowledge text in the agent Prompt & Goals step.
  • A linked published knowledge base created under Knowledge base (/knowledge-base in the app).
1

Detect an information need

The runtime identifies that the current turn requires factual grounding or tenant-specific data.
2

Query the approved knowledge source

Only the configured knowledge base or allowed retrieval surfaces are queried.
3

Return a grounded response

The model uses the retrieved evidence to answer directly and concisely.
4

Log the outcome

Retrieval-backed behavior can be inspected through transcript and trace data for QA and continuous improvement.
Public OpenAPI does not yet generate the full knowledge-bases family. Route ownership and scopes are live in the product (knowledge_bases:read / knowledge_bases:write). Use ../guides/knowledge-bases until generated coverage is complete.

​
Knowledge quality checklist

  • Keep sources scoped by tenant/domain and publish state.
  • Remove stale or duplicate documents that compete for the same intent.
  • Include versioned process documents for policy-sensitive answers.
  • Audit retrieval-backed transcripts to identify missing knowledge coverage.
Do not give the runtime unrestricted access to large, noisy document sets. Retrieval quality depends on disciplined knowledge curation and publication review.