Document Section Finder

Upload PDF, DOCX, or TXT files, ask a natural-language question, and get back the most relevant paragraphs with their page numbers. Entirely local — no API keys required.

Built with Streamlit, sentence-transformers, ChromaDB, and a free local LLM.

---

Features

Core search

• Multi-format ingestion — PDF, DOCX, and TXT in a single upload step.
• Paragraph-level search — documents are split into paragraphs, each embedded for semantic search.
• Page number tracking — PDF page numbers are detected and displayed in results.
• Hybrid search (BM25 + semantic) — combines keyword matching with dense embeddings for better recall. Weight adjustable.
• Cross-encoder re-ranking — re-ranks top results with cross-encoder/ms-marco-MiniLM-L-6-v2 for higher precision.
• Multi-model embedding selector — choose from 5 free sentence-transformer models (small/fast to high-quality).
• Tunable relevance — adjust results count and minimum similarity threshold.
• Highlighted match snippets — query-relevant sentences extracted with sentence-level embeddings; query terms highlighted in <mark> tags.
• One-click copy — copy matched paragraph references to clipboard.

Advanced

• RAG (Retrieval-Augmented Generation) — toggle to generate a natural-language answer from a free local LLM (google/flan-t5-large) with source citations.
• Persistent vector index — embeddings saved to disk via ChromaDB, surviving app restarts. Save / load with one click.
• OCR for scanned PDFs — automatically falls back to pytesseract when PyPDF2 extracts no text.
• PDF report export — download search results as a formatted PDF document.
• Keyboard shortcuts — Ctrl+Enter to search, Esc to clear input.

Privacy

• 100% local — no documents leave your machine; no API keys required.
• No telemetry — ChromaDB telemetry is disabled.

---

Architecture

app.py                  Streamlit UI (tabs, sidebar, keyboard shortcuts)
├── document_processor  File parsing, OCR, paragraph splitting, page tracking
├── semantic_search     BM25 + semantic hybrid search, cross-encoder re-ranker
├── vector_index        ChromaDB persistent vector storage
├── rag_engine          Free local LLM for AI answer generation
├── export_utils        PDF report generation
└── styles.css          UI styling (light theme)

Workflow

1. Upload — drop PDF/DOCX/TXT files into the UI.
2. Parse — text extracted per file type; OCR fallback for scanned PDFs.
3. Split — text split into paragraphs; PDF page markers tracked.
4. Embed — each paragraph encoded into a vector using the selected embedding model.
5. Search — query embedded and ranked via hybrid BM25 + cosine similarity; optionally re-ranked by a cross-encoder.
6. Snippet extraction — for each result, the most query-relevant sentences are extracted and highlighted.
7. RAG (optional) — the top paragraphs are fed to a local LLM for a generated answer with citations.
8. Render — results shown as expandable cards with paragraph number, page, source, and highlighted snippet.

---

Installation

Requires Python 3.9+.

git clone https://github.com/Abdelrahman-Atef-Elsayed/document-section-finder.git
cd document-section-finder

python -m venv .venv
# Windows
.venv\Scripts\activate
# macOS / Linux
source .venv/bin/activate

pip install -r requirements.txt

Optional dependencies

Feature	Package	Download size
Hybrid search (BM25)	rank-bm25	small
Cross-encoder re-ranker	sentence-transformers (built-in)	~80 MB
RAG (AI answers)	transformers	~800 MB (flan-t5-large)
Persistent index	chromadb	small
OCR for scanned PDFs	pytesseract + Pillow	small (+ Tesseract system binary)
PDF export	fpdf2	small

All models download automatically on first use and cache locally.

---

Usage

streamlit run app.py

Streamlit opens http://localhost:8501 in your browser.

1. Upload tab — drop files and click Process Files.
2. Search tab — type a query and click Search (or Ctrl+Enter).
3. Browse results with paragraph numbers, page numbers, and highlighted snippets.
4. Toggle Hybrid search, Cross-encoder, or AI answer in the sidebar.
5. Click Export Results as PDF to download a report.

Sidebar settings

Setting	Description
Embedding model	Choose from 5 sentence-transformers models
Hybrid search	BM25 keyword + semantic cosine fusion
Cross-encoder	Precision re-ranker on top results
AI answer (RAG)	Local LLM generates answer with citations

Persistent index

After processing documents, click Save to Index to persist embeddings to disk. On subsequent launches, click Load Saved Index to restore without re-uploading.

Example queries

• What are the requirements for user authentication?
• Data retention and privacy policies
• System architecture and design overview

---

Tech Stack

Layer	Technology
UI	Streamlit
Embedding models	paraphrase-MiniLM-L6-v2, all-MiniLM-L6-v2, multi-qa-MiniLM-L6-dot-v1, all-mpnet-base-v2, BAAI/bge-small-en-v1.5
Cross-encoder	cross-encoder/ms-marco-MiniLM-L-6-v2
RAG LLM	google/flan-t5-large (via Hugging Face Transformers)
Vector index	ChromaDB (persistent, local)
Tensor backend	PyTorch (CPU)
Keyword search	BM25 (rank-bm25)
Data handling	pandas
PDF parsing	PyPDF2 + pytesseract (OCR fallback)
DOCX parsing	python-docx
PDF export	fpdf2
Language	Python 3.9+

---

Project Structure

.
├── app.py                  # Streamlit UI entry point
├── document_processor.py   # File parsing, OCR, paragraph splitting
├── semantic_search.py      # Hybrid search, cross-encoder, multi-model
├── vector_index.py         # ChromaDB persistent vector storage
├── rag_engine.py           # Local LLM for AI answer generation
├── export_utils.py         # PDF report export
├── styles.css              # UI styling
├── requirements.txt        # Python dependencies
├── LICENSE                 # MIT
└── README.md

---

License

Released under the MIT License.