This project is only for educational purpose, it aims to teach myself about the fundemental workflow of RAG (Retrieval-Augmented Generation). The app is powered by FastAPI with Jinja2 template engine. Besides that, for autocomplete system, the trie data structured is also utilized (using submodule named trie).
- Autocomplete Functionality: Utilizes a trie data structure for efficient autocomplete suggestions based on LLM generated questions.
- RAG System: Leverages the concept of Retrieval-Augmented Generation for enhancing information retrieval and response generation (In progress).
- LLM Integration: Uses a Large Language Model to analyze data, identify trends, and generate relevant questions. [Looking options for utilizing open-source LLMs with faster inference methods (tvm, llama.cpp, wasm)].
- LLM Used:
-- Mistral-7B API provided by together.ai
-- GPT-3.5 powered by OpenAI
Clone:
https://github.com/author31/funny-rag.git
Install packages
pip install -r requirements.txt
Run:
uvicorn main:app --port 8000 --host 0.0.0.0 --reload
- Crawl the corresponding post linked urls to: -- To generate comprehensive questions for clustered posts. -- To generate insightful answer.
- Build connections between LLMs and data source
- Deploy to AWS, only using free-tier service.