Context-Aware Email Assistant (RAG over Gmail .mbox + Fine-tuned Gemma)

A local-first Retrieval-Augmented Generation (RAG) system that answers questions over Gmail archives exported via Google Takeout (.mbox). It parses/cleans emails, builds a vector index, retrieves relevant context, and generates grounded answers using local LLM inference. Includes an optional Gemma fine-tuning notebook (QLoRA).

Problem

Keyword search isn’t enough for large mailboxes. This project enables semantic search + QA over historical emails while keeping data local.

Architecture

Takeout (.mbox) → Parse/Clean → Chunk → Embed → ChromaDB → Retrieve (top-k) → LLM → Answer

Tech Stack

• Python, Jupyter
• LangChain (RAG)
• ChromaDB (vector store)
• GPT4All embeddings (embeddings)
• Local inference (GPT4All / llama.cpp)
• BeautifulSoup (HTML cleaning)
• Optional fine-tuning: Transformers + TRL + PEFT (LoRA) + bitsandbytes (4-bit)

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Results

• Enables semantic search + question answering over Gmail Takeout exports while keeping data local.
• Retrieves top-k relevant email chunks and generates grounded answers with citations to source messages.
• Designed to avoid committing sensitive data (Takeout/.mbox ignored via .gitignore).