Local RAG for your files and inbox, no cloud required

What it does

A handful of Python scripts turn Ollama local LLMs into a retrieval-augmented generator for your own documents and emails. You feed it PDFs, .txt, or JSON via upload.py, or let collect_emails.py download your inbox; the chat scripts then retrieve context and answer questions while keeping everything on your machine. It is essentially friendly glue code around Ollama and an embedding model.

The interesting bit

The README is mostly setup steps and YouTube links, yet the codebase itself is admirably focused: it offers both a query-rewrite path (localrag.py) for vague questions and a blunt no-rewrite path (localrag_no_rewrite.py), plus a true conversation loop with history. The email RAG feature is a nice pivot from the usual “chat with your PDF” demo.

Key highlights

• Runs entirely offline using local Ollama models and the mxbai-embed-large embedding model.
• Ingests PDFs, plain text, and JSON through upload.py; email history via collect_emails.py.
• Query rewrite mode attempts to improve retrieval when your question is ambiguous.
• CLI model picker (--model mistral, etc.) with llama3 as the default.
• Maintains conversation history across turns instead of treating each question in isolation.

Caveats

• The README never names the underlying vector database, so how embeddings are stored and retrieved is unclear.
• Email access requires stuffing your credentials into a .env file and generating app-specific passwords.
• The project is tightly coupled to Ollama; swapping in another local inference backend would require rewriting the scripts.

Verdict

Worth a look if you want a minimal, hackable local RAG starter. Look elsewhere if you need enterprise scalability or detailed architecture docs.