Built using FastAPI, LilLisa Server is an AI question-and-answer program that enables users to ask natural language questions and receive insightful answers based on their personalized data.
LilLisa Server uses LanceDB as the vector database, providing blazingly fast retrieval times and LlamaIndex to handle complex queries.
LilLisa Server is able to grow its knowledge base and get better at answering user questions as more documentation is added and conversations are stored.
- Clone this project using this command:
- Navigate to lil-lisa folder
- In the terminal, run "make condaenv". This will create a conda environment and install all necessary packages
- Select 'lillisa-server' as python interpreter
IMPORTANT: If using VS Code, start main.py by using "Python Debugger: Debug using launch.json" Integrate with Slack, FastHTML, or another application that handles user input. Run one of the above along with LilLisa_Server conccurently.
Slash commands are encrypoted and can only be used by admins specified in Slack.
Methods free to use are:
/invoke
Uses a session ID to retrieve past conversation history. Based on a query, it searches relevant documents in the knowledge base and retrieves multiple fewshot examples from the QA pairs database to help synthesis of a formatted answer. Queries are handled differently and depend on whether an expert is answering or not. ReACT agent handles the use of information given to intellignetly craft an answer.
/record_endorsement
Records an endorsemewnt, usually given when an answer is correct, by either a "user" or "expert". This is helpful when admins call the 'get_conversations' method and use it to create more golden QA pairs.
The project is not currently open for contributions.
- Docker container
- Python 3.11.9
- RAM: 1.0 GB
- Size of Docker container: 11.2 GB
For assistance with deploying to AWS Lambda, refer to this blog:
Reach out to us if you have questions:
- Carlos Escobar (Slack: @Carlos Escobar, Email: [email protected])
- Dhar Rawal (Slack: @Dhar Rawal, Email: [email protected])
- Unsh Rawal (Slack: @Unsh Rawal, Email: [email protected])
- Carlos Escobar
- Dhar Rawal
- Unsh Rawal
- Nico Guyot
This project is currently closed source
Under active development