An AI agent that leverages Pydantic AI documentation to build AI agents

Multi-agent architecture built using Pydantic AI, LangGraph, and LangChain

Agents that coordinate reasoning (with qwen3), tool use, code writing (with qwen2.5 coder) and output formatting

Agentic RAG pipeline integrating Supabase for indexing and retrieval, with a script that crawls the entire Pydantic AI documentation

Observability and tracing using Langfuse

Streamlit frontend to interact with the agent

Deployed locally using Docker for frontend and database, and LM studio (with ollama) to serve models

Front-end: Streamlit Models: LLMs and embeddings Database: Supabase

Pydantic AI (for agents)

Langchain (for summarizing content with LLM and interacting with embeddings models)

LangGraph

Langfuse (for observability and tracing)

Supabase

Crawl4AI (for crawling documentation content)

Streamlit

Clone this repo

Create and activate a virtual environment

Run pip install -r requirements.txt

Spin up supabase locally or create a project on the cloud

Spin up langfuse locally or create a project on the cloud

Rename .env.example to .env and add the necessary variables

Run python crawl_pydantic_ai_docs.py to crawl and index the documentation

Run Streamlit run streamlit_ui.py to start talking with the agent

I tested everything on a local setup using LM studio to serve Qwen3 8B for reasoning, Qwen2.5 Coder 7B Instruct for code generation, and Text Embedding Nomic Embed Text v1.5 for embeddings generation. Along with a docker deployed instance of supabase and langfuse. Here's why I changed from Llama 3.2 3B Instruct to Qwen:

Qwen3 8B was giving me much better results with reasoning, especially after I fixed the bug with <think> tags for tooling, I wanted to test Deepseek R1 Distill Llama 8B but Qwen is already giving really good results so it became unnecessary to explore.

Upgrading from QnA to code generation required using an LLM that is specifically trained for code generation and Qwen2.5 is a very powerful model at that (maybe not as powerful as Devstral but the 7B version is definitely the best my machine could handle, also it's already giving me very good results so I felt no need to explore other options). I tried using Qwen3 for code generation too but the model was getting stuck in an infinite loop of tooling calls.

As for embeddings, I tried Bge Large En v1.5 but it kept crashing due to local setup limitations.

Once the UI is up and running, streamlit should automatically open a browser window with the agent, or at least print the URL in the terminal

Talk to the agent and ask it to code an agent

After the model responds, you can check traces on langfuse dashboard