FastRAG — Data flow Diagram

Docs and code are split by heading or AST symbol; prose is translated to hi, bn, ta, te, mr and cached under eval/translations
Unanswerable candidates whose top-1 score reaches the answerable lower quartile are dropped as leaks
Calibration fits the abstention, CRAG, cache-distance and off-topic gates from the calibration split and cache pairs
Not drawn: the eval-set builder reads top-1 scores back from the new collection, and the API loads calibration gates and the centroid at startup

Questions pass guardrails, the caches, hybrid retrieval, rerank and CRAG before the LLM sees any context
Answers stream to the browser sentence by sentence; the final event carries citations and the full trace

Query text and retrieved passages go to Jina and the LLM provider; audio goes to Sarvam
Langfuse receives spans with content hashed unless FASTRAG_TRACE_RAW_CONTENT is set
Every stage also emits Langfuse spans and Prometheus metrics (/metrics, scraped every 15s in compose); not drawn