DocuChat

A Retrieval-Augmented Generation (RAG) app for querying your own PDFs/notes, running entirely on local compute except for the final answer generation call.

Pipeline: Sentence-Transformer embeddings → local FAISS vector search → cross-encoder reranking → grounded answer generation via the Gemini API → Streamlit UI with source-passage citations.

Setup

python3 -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env
# then edit .env and paste your Gemini API key
# (get a free one at https://aistudio.google.com/apikey)

Usage

streamlit run app.py

1. In the sidebar, upload one or more .pdf, .txt, or .md files.
2. Click Build / Rebuild Index — this chunks the text, embeds it with Sentence-Transformers, and builds a local FAISS index.
3. Ask questions in the chat box. Each answer includes an expandable Sources section showing which passages (and page numbers) were used.

How it works

1. Ingest (src/ingest.py): documents are loaded, split into overlapping chunks with LangChain's RecursiveCharacterTextSplitter, embedded with a Sentence-Transformer model (all-MiniLM-L6-v2 by default), and stored in a local FAISS IndexFlatIP index (cosine similarity via normalized vectors).
2. Retrieve (src/retriever.py): the query is embedded and FAISS returns the top-N candidate chunks. A cross-encoder (cross-encoder/ms-marco-MiniLM-L-6-v2) then rescoring the (query, chunk) pairs directly, which is more accurate than embedding similarity alone, and the top-K are kept.
3. Generate (src/generator.py): the reranked passages are inserted into a grounded prompt and sent to the Gemini API, which is instructed to answer only from the given context and cite sources inline.
4. UI (app.py): Streamlit handles uploads, indexing, chat history, and rendering of the cited source passages.

Rebuilding the index

Re-run indexing any time you add or change documents in data/raw_docs/ — either via the sidebar button or:

python -m src.ingest

Notes

• Everything except the final Gemini call runs locally — no cloud vector DB, no external embedding API.
• Swap EMBEDDING_MODEL, CROSS_ENCODER_MODEL, GEMINI_MODEL, chunk sizes, and top-k values in .env without touching code.