OpenTranscribe — Search & Indexing

OpenTranscribe — Search & Indexing A workflow diagram generated by Archify. 01 / API / Web 02 / OpenSearch 03 / Indexing (write path) Query Filter + return User · search box · API / Web › Query User search box Search API · parse query + filters · API / Web › Query Search API parse query + filters Filter · quarantine + snippet redaction · API / Web › Filter + return Filter quarantine + snippet redaction Results · highlighted snippets · API / Web Results highlighted snippets RRF Fusion · BM25 + kNN, one pipeline · OpenSearch › Query RRF Fusion BM25 + kNN, one pipeline OpenSearch · hybrid ranked hits · OpenSearch › Filter + return OpenSearch hybrid ranked hits Celery · post-transcription · Indexing (write path) › Query Celery post-transcription Chunker · chunk + embed transcript · Indexing (write path) › Query Chunker chunk + embed transcript query + filters ranked hits index chunks + whole-doc Legend User UI Agent logic Policy Tool action Context / trace External system

One pipeline, two signals

  • • cosinesimil returns (1+cosine)/2, not raw cosine — every kNN read converts back
  • • RRF combines lexical (BM25) and semantic (kNN) ranks server-side

Filtering happens after ranking

  • • Quarantined files are dropped from hits before the client ever sees them
  • • Snippet redaction masks PII in highlighted text, independent of ranking

Write path is async

  • • Chunking + embedding runs once per transcript, dispatched right after finalize
  • • Legacy whole-doc index kept alongside chunk-level index for back-compat