LLM CHAT ASSISTENT

A Python library and web app for an LLM-based chatbot with many capabilities.

Features:

• Privacy: all messages and uploaded attachments are encrypted so that no-one can listen in on your conversation
• Knowledge: access to documentation
• Continuous conversation: conversations are summarized when they become too long to fit into the prompt
• Tool use:
  • Code execution: ability to execute Python and R code
  • Google Scholar search: ability to search for articles on Google Scholar
  • Attachments: ability to read attachments
  • Download: ability to download pages and files as attachments

Sigmund is not a large language model itself. Rather it uses third-party models. Currently, models from OpenAI, Anthropic, and Mistral are supported. API keys from these respective providers are required.

By default, Sigmund is configured to act as an assistant for OpenSesame, a program for creating psychology/ cognitive-neuroscience experiments. However, the software can easily be reconfigured for a different purpose.

Configuration

See heymans/config.py for configuration instructions.

Dependencies

For Python dependencies, see pyproject.toml. In addition to these, pandoc is required for the ability to read attachments, and a local redis server needs to run for persistent data between sessions.

Running (development)

Download the source code, and in the folder of the source code execute the following:

# Specify API keys for model providers. Even when using Anthropic (Claude) or
# Mistral, an OpenAI key is provided when document search is enabled
export OPENAI_API_KEY = 'your key here'
export ANTHROPIC_API_KEY = 'your key here'
export MISTRAL_API_KEY = 'your key here'
pip install .               # install dependencies
python index_library.py     # build library (documentation) index
python app.py               # start the app

Next, access the app (by default) through:

https://127.0.0.1:5000/

Running (production)

In production, the server is generally not run by directly calling the app. There are many ways to run a Flask app in production. One way is to use gunicorn to start the app, and then use an nginx web server as a proxy that reroutes requests to the app. When taking this route, make sure to set up nginx with a large client_max_body_size (to allow attachment uploading) and disable proxy_cache and proxy_buffering (to allow status messages to be streamed while Sigmund is answering).