LAMBDA (LocalAutoMailBoxDraftAssistant)

A local AI-powered email automation system that learns from your email style and creates draft responses for every unread email in your (Gmail) inbox.

demo.mp4

Save time replying to emails -- JUST Open the draft and Send
Keep all data and model to yourself -- No cloud, no tracking
Something everyone can use everyday -- Set up in less than 10 minutes and forget about it

Features

Automatically extracts your email communication patterns from Gmail
Fine-tunes LLaMA models on your email style using LoRA
Creates AI-generated draft responses for unread emails
Supports both Apple Silicon (MLX) and CUDA GPUs

Requirements

Python 3.10+
Gmail Account with API access
For Mac: MLX compatible device (M1/M2/M3/M4/...)
For others: CUDA compatible GPU

Installation

Clone the repository:

git clone https://github.com/zycyc/LAMBDA.git
cd LAMBDA

Install dependencies:

For Mac (Apple Silicon) and Linux:

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

For Windows:

python -m venv .venv
Set-ExecutionPolicy Unrestricted -Scope Process
.venv\Scripts\activate
pip install -r requirements.txt

Note: If you encounter any issues with PyTorch installation:

For Windows, you can manually install PyTorch with CUDA using:

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

For Linux, you can manually install PyTorch using:
```
pip3 install torch torchvision torchaudio
```

Enable Gmail API and Setup Credentials (only takes about 5 minutes):
1. Go to Google Cloud Console
2. Create a new project:
  - Click on the project dropdown at the top of the page
  - Click "New Project"
  - Enter a project name (e.g., "LAMBDA Email Assistant")
  - Click "Create"
  - Wait for the project to be created and click "SELECT PROJECT"
3. Enable Gmail API:
  - Click on the hamburger menu (☰) in the top-left corner
  - Navigate to "APIs & Services" > "Library"
  - Search for "Gmail API"
  - Click on "Gmail API" in the results
  - Click "Enable"
4. Configure OAuth consent screen:
  - Go to "APIs & Services" > "OAuth consent screen"
  - Select "External" user type (unless you're in an organization)
  - Click "Create"
  - Fill in the required fields:
    - App name: "LAMBDA Email Assistant"
    - User support email: your email
    - Developer contact information: your email
  - Click "Save and Continue"
  - On "Scopes" page
    - Click "ADD OR REMOVE SCOPES"
    - Search for "Gmail API" and select "Read, compose, and send emails from your Gmail account"
    - Click "UPDATE" and then "SAVE AND CONTINUE"
  - On "Test users" page, click "ADD USER" and enter your email, then click "SAVE AND CONTINUE"
  - Click "Back to Dashboard"
5. Create OAuth 2.0 credentials:
  - Go to "APIs & Services" > "Credentials"
  - Click "Create Credentials" at the top
  - Select "OAuth client ID"
  - Choose "Desktop app" as the application type
  - Name it "LAMBDA Desktop Client"
  - Click "Create"
6. Download credentials:
  - In the popup that appears, click "Download" (or download from the credentials page)
  - Rename the downloaded file to credentials.json
  - Move the file to your LAMBDA project root directory
Note: When you first run the application, it will open a browser window asking you to authorize the application. This is normal and only needs to be done once.

Usage

Run the LAMBDA interface:

python lambda.py

or run the bot in background once you're satisfied with everything:

nohup python lambda_bot.py &> lambda_bot.log &

The interactive menu will guide you through the following options:

Create/Update training dataset
Train model
Run email bot
Run complete workflow (all steps)
Exit

On first run, you'll be prompted to enter your name if not already configured. This name will be used by the AI to personalize responses.

The bot will:

Check for unread emails every 5 minutes (adjustable in config.py)
Generate responses using your fine-tuned model
Create draft responses in Gmail
Skip threads that already have drafts

Project Structure

LAMBDA/
├── README.md
├── requirements.txt
├── blacklist.txt          # Sender blacklist to ignore
├── config.py              # Configuration settings
├── lambda.py              # Main interface script
├── create_dataset.py      # Gmail data extraction
├── train_model.py         # Model fine-tuning
├── lambda_bot.py          # Email bot implementation
└── gmail_utils.py         # Gmail API utilities

Configuration

Edit config.py to customize:

Model selection
Training hyperparameters
Email settings
Response templates

Email Filtering

LAMBDA provides multiple ways to filter which emails get processed:

1. Gmail Labels (Automatic)

By default, LAMBDA skips emails with the following Gmail labels:

CATEGORY_SOCIAL (social media notifications)
CATEGORY_UPDATES (automatic updates)
CATEGORY_FORUMS (forum posts)
CATEGORY_PROMOTIONS (marketing emails)

2. Sender Blacklist (Manual)

You can manually specify which senders to ignore by adding their email addresses to blacklist.txt:

Open blacklist.txt in any text editor
Add one email address per line
You can use partial matches - they will match anywhere in the sender's email
Lines starting with # are treated as comments

Example blacklist.txt:

no-reply@
newsletter@
notifications@
marketing@

This would skip any emails from addresses containing these strings (e.g., no-reply@company.com, marketing@example.com)

Q&A

Can I run this on X computer with a Y GPU?
- Non-M-chip Macs are not supported. Otherwise, as long as it comes with a M-Chip or a CUDA-compatible GPU, it should work.
Do I have enough GPU power to run this? Can I run smaller/quicker models?
- It reallydepends. In general, I'd recommend asking any LLMs for a proper setup of hyperparameters that you'd want to put in config.py. For Mac users, mlx docs offers useful instructions (including how to convert the default LoRA model to QLoRA format).
Will this work with my Outlook email?
- No, this is a Gmail-only tool for now. I got some example scripts for Outlook, but they are not fully tested & integrated yet. The key difficulty is that there're many Gmail-specific APIs that have already been built into the current version.
Anything else?
- Feel free to file an issue on Github!

TODO

CUDA QLoRA support
Windows support (now only tested on Mac)
Filter out spam ads emails by using replied rate (or is there some API for this?)
A configurable list of senders to ignore (e.g., no-reply@example.com)
Shortcuts support for voice-input-based rewriting (whisper)
RAG support for personal knowledge base
Outlook Exchange support
Labeling emails for priorities / categories
...

License

MIT License - see LICENSE file for details

Acknowledgements

MLX team for bringing LLMs to M-chip Macs with increasingly improved APIs
LLMMe for initiating this project (I rewrote and streamlined the code to be more user-friendly)