Khatwanigaurav/Doc_QA_Agent
0
PolicyQA — Document QA Agent
Upload a policy document or report (PDF or plain text). Ask questions in natural language. Get answers cited to exact sections.
Pipeline
Upload PDF / paste text
│
▼
pymupdf extracts text
│
▼
Section-aware parser (detects headings → builds §path per section)
│
▼
Chunk each section (600 chars / 100 overlap, heading prepended)
│
├── BGE-base-en-v1.5 → FAISS (dense, top-20)
├── BM25Okapi (sparse, top-20)
└── RRF fusion k=60 (merged, top-20)
│
▼
BGE-reranker-large (cross-encoder, top-5)
│
▼
flan-t5-xl (generator, ~3B)
│
▼
flan-ul2 (judge, ~20B) score 1–10
│
score < 6 → escalate
score ≥ 6 → cited answer