OpenTranscribe — RAG Chat Retrieval

OpenTranscribe — RAG Chat Retrieval A sequence diagram generated by Archify. ask question resolve scope + plan kNN + keyword query ranked chunks/digests mask_chunks / mask_digests unmasked (local) or masked (remote) draft answer answer + citations (SSE) Plan Retrieve + guard Generate + answer Chat UI · browser · Sequence participant Chat UI browser Chat Router · plan + scope · Sequence participant Chat Router plan + scope Retrieval · chunks + digests · Sequence participant Retrieval chunks + digests OpenSearch · hybrid/neural search · Sequence participant OpenSearch hybrid/neural search Redactor · mask if remote · Sequence participant Redactor mask if remote LLM · local or remote · Sequence participant LLM local or remote Output Mask · display-layer mask · Sequence participant Output Mask display-layer mask Legend request return security default message

Two independent masks

  • • Egress mask (redactor): keyed on provider locality, decides what the LLM sees
  • • Output mask (output_redactor): display-layer, independent of the egress policy

Fail-closed

  • • Any ambiguity in provider classification reads as remote
  • • An admin's force_redact_before_llm lock always wins the local exemption

Two retrieval units

  • • mask_chunks addresses text by time range
  • • mask_digests addresses text by provenance (segment_ids) — not interchangeable