RAG-Powered Q&A on Personal Docs (Local, Free)

This is a tiny, fully local Retrieval-Augmented QA system you can run in a few minutes. It demonstrates the full RAG loop using free components only.

High-level flow

1. Ingest: Read PDFs/Markdown/TXT → split into overlapping chunks (to keep context).
2. Embed: Convert chunks to vectors with a lightweight model (all-MiniLM-L6-v2).
3. Index: Store vectors in FAISS (fast vector search) + parquet metadata.
4. Query: User question → embed → retrieve top-k similar chunks.
5. Answer: Show evidence chunks and a simple extractive "answer" assembled from top hits.

Abstractive answering (a generated summary) can be plugged in later with a small local model.

Quick Start

python -m venv .venv
# macOS/Linux
source .venv/bin/activate
# Windows (PowerShell)
# .venv\Scripts\Activate.ps1

pip install -r requirements.txt
streamlit run app.py

Then in the UI:

1. Upload 1+ documents (or place them in data/source/).
2. Click Build / Refresh Index.
3. Ask a question and inspect the Evidence + Answer sections.

Key design choices

• Cosine similarity via Inner Product: Normalize embeddings and use FAISS's IndexFlatIP.
• Chunking: Default chunk_size=800, overlap=120. Adjust for your content density.
• Local-first: No paid APIs. Everything runs on CPU.

Files

rag-docs/
├─ app.py                 # Streamlit UI: upload → build index → query
├─ ingest.py              # Ingestion pipeline (read/clean → split → embed → index)
├─ rag_core/
│  ├─ splitter.py         # Chunking utilities (with comments)
│  ├─ embedder.py         # Wrapper around SentenceTransformer (normalized vectors)
│  ├─ index.py            # FAISS + metadata persistence
│  └─ answer.py           # Naive extractive answer composer
├─ data/source/           # Put your PDFs/MD/TXT here (UI saves uploads here too)
└─ data/store/            # FAISS index + Parquet metadata live here

Next steps (nice v2s)

• Add abstractive summarization (e.g., google/flan-t5-small) over retrieved chunks.
• Highlight/cite exact phrases that matched the query.
• Add OCR for scanned PDFs (Tesseract) if needed.