This is a Streamlit-based application that allows users to upload various document types and ask natural language questions about their content. The app extracts text, normalizes it, handles search using a hybrid retrieval setup, and answers questions using an LLM with precise source citations.
- Multi-Format Ingestion: Supports uploading PDF (scanned/digital), DOCX, PPTX, TXT, MD, and JSON files.
- Advanced Text Processing: Normalizes text by removing layout noise, fixing line-break hyphenations, and clearing hidden control characters.
- Hybrid Retrieval Pipeline: Combines dense semantic search (FAISS) with exact keyword matching (BM25) using Reciprocal Rank Fusion (RRF) for highly accurate document retrieval.
- Two-Stage Reranking: Filters retrieved text using a Cross-Encoder to prioritize the most contextually relevant chunks.
- Source Confidence Filter: Blocks irrelevant queries from reaching the LLM if the document corpus does not contain the answer.
- Citation-Backed Answers: Formats responses with interactive, superscript inline footnotes that link directly to specific page numbers, slides, or text blocks.
- Hallucination Safeguards: Cross-checks generated answers against raw source documents using a self-reflective validation layer.
- Semantic Caching & Memory: Intercepts repeating queries instantly using a vector cache (saving API calls) and maintains a sliding conversation history for multi-turn chat.
- Frontend: Streamlit
- Core Languages: Python
- Parsing & OCR: PyMuPDF, python-docx, python-pptx, EasyOCR, Pillow
- Vector Search & Indexing: FAISS (Dense) & Rank-BM25 (Sparse)
- NLP & Reranking: Sentence Transformers (Embedding & Cross-Encoder)
- Inference Interface: HuggingFace Hub Inference Client (Qwen/Qwen2.5-7B-Instruct)
├── app.py # Streamlit user interface & state coordinator
├── requirements.txt # Project dependencies
└── scripts/
├── __init__.py # Package initialization marker
├── document_parser.py # Path-aware multimodal document reader
├── cleaner.py # In-place unicode and spacing normalizer
├── chunker.py # Whitespace-safe sliding window text splitter
├── embedding.py # Sentence Transformer vector generation wrapper
├── vector_db.py # Hybrid retriever, RRF, reranker, and confidence filter
├── qa_model.py # Structured JSON response generator and self-audit layer
├── orchestrator.py # Session memory, semantic cache, and logging tracker
└── evaluator.py # Programmatic LLM-as-a-Judge benchmarking suite
- Install Dependencies:
pip install uv
uv sync
- Configure Environment Variables:
Create a
.envfile in the root directory and add your Hugging Face API token:
HF_TOKEN=your_huggingface_token_here
- Run the Application:
streamlit run app.py