RESEARCH_ONLY · PAPER_TRADING_ONLY · PUBLIC_MARKET_DATA_ONLY · USD_ONLY · NO_REAL_ORDER_CAPABILITY
OV
OV Research Desktop
Codebase-derived project showcase

RAG and Evidence

How the hybrid retrieval system finds and governs evidence — verified against src/retrieval/hybrid.py and src/chunking.

Retrieval pipeline

CODE-DERIVED VISUALISATION
  1. 1Document ingestion (src/ingestion)
  2. 2Chunk generation (src/chunking)
  3. 3Embeddings (src/embeddings)
  4. 4Qdrant collection storage (vector_store/, requirements.txt: qdrant-client)
  5. 5BM25 always runs; dense vector search when available (src/retrieval/hybrid.py)
  6. 6Source-approval filter (only OV-0001..OV-0004 chunks are retrievable)
  7. 7Candidate-aware evidence ranking
  8. 8Semantic validation + contradiction resolution (src/extraction)
  9. 9Governed evidence summary presented to the research pipeline

Synthetic worked example

CODE-DERIVED VISUALISATION
CODE-DERIVED VISUALISATION
  1. 1Runtime query: "What confirms a 25% counter-trend scalp setup?" (synthetic, not a real user submission)
  2. 2Query restricted to the approved source filter (OV-0001..OV-0004 only)
  3. 3Hybrid BM25 + vector retrieval returns candidate chunks
  4. 4Contradiction check against src/validation indexes
  5. 5Evidence summary: 2-3 short excerpts with source IDs, not a full transcript
  6. 6Governed result handed to the strategy/decision layer

This example uses a synthetic question for illustration only — it is not a record of a real query ever made against this system.