Project 05 / 10 · RAG & NLP

DocuQuery AI Assistant

Production RAG system for natural-language QA over PDFs — hybrid retrieval with LangGraph multi-intent orchestration. Submitted to IEEE Access.

01 / Context

The problem

Document questions vary between conceptual paraphrases, noisy OCR text, multilingual passages, and exact identifiers. A single fixed retrieval blend can silently favor one query type while weakening another.

02 / System workflow

From input to output

  1. 01PDF upload
  2. 02Hierarchical chunks
  3. 03Three retrievers
  4. 04Rerank + diversify
  5. 05LangGraph intent
  6. 06Grounded answer

03 / Architecture

Inside the system

A1

pypdf extraction with OCR fallback feeds a two-scale hierarchical chunker that preserves semantic boundaries.

A2

FAISS dense search, BM25, and TF-IDF candidates are fused, cross-encoder reranked, and diversified with MMR; Pinecone is available as an optional cloud index.

A3

LangGraph routes QA, summarisation, comparison, and refinement before Gemini generates an answer with confidence and evidence chunks.

04 / Engineering decisions

What shaped the build

Decision 01

Hybrid by evidence

Lexical and dense retrieval are benchmarked separately so fusion strengths and failure modes remain visible.

Decision 02

CPU-aware embeddings

MiniLM remains the default for accessible deployment, while the benchmark documents the quality lift from a larger MPNet embedder.

Decision 03

Reproducible evaluation

The benchmark covers 1,000 passages and 100 queries across four scenarios with McNemar tests and bootstrap confidence intervals.

05 / Evidence

Measured, not implied

1.00

Clean Recall@5

Hybrid retrieval on clean paraphrase and multilingual scenarios.

+73pp

OCR robustness

Gain over dense-only retrieval with 2% character corruption.

47ms

P95 end to end

Local query-to-extractive-answer latency over 1,000 chunks.

06 / Delivery

What shipped

  • 01The Gradio application accepts multiple PDFs, supports follow-up questions and document comparison, and exposes extracted text and chunk previews.
  • 02Conversations export to TXT, PDF, or DOCX.
  • 03The application is deployed as a Hugging Face Space with configurable Gemini and Pinecone integrations.

Constraint documented

Pure BM25 beats the fixed hybrid on rare identifier lookups because dense scores dilute exact-match evidence. That documented failure motivates query-aware fusion gating.