M10 · RAG · vector DB · reranking

RAG and retrieval

Hybrid retrieval with metadata filters, reranking, citation and permission checks.

$ python -m fde_toolkit rag --query "..."

Workflow

the order an FDE actually runs it in

  1. 1Parse and structure the document, then chunk semantically.
  2. 2Embed and index with tenant, role and date metadata.
  3. 3Retrieve with hybrid dense plus keyword scoring.
  4. 4Filter by permission and metadata before reranking.
  5. 5Select evidence, generate, and cite every claim.

Function output

deterministic trace

query='EDD refresh policy'  strategy=sentence  top_k=5
dense 0.456 · keyword 0.570 -> hybrid 0.6(0.456) + 0.4(0.570) = 0.502
chunk #4 selected, 1 citation attached, 2 chunks dropped by tenant filter