Day 29 – Building my own patent chat with a graph RAG

🔥

Today I wanted to talk to the data that now lives inside my Graph RAG and my Supabase vector database. So I rebuilt the chat pipeline to really understand what is happening behind the scenes.

And just like yesterday… this Graph RAG truly feels like a monster.

What came out of it is something I honestly find fascinating.

I built my own personal patent assistant.

I can now chat with my stored patent data. The system understands the meaning of my questions and automatically searches across all tables for the right information. The agent then connects everything together into a clear answer and, thanks to memory, it does not forget the context of our conversation.

So instead of digging through endless lists, I can simply ask and get exactly what I need.

There is a reason why this pipeline has so many nodes.

My AI agent needs two different tools for each of my three tables: metadata, claims and prior art.

One set of tools helps the agent understand the question and find the right documents.The other set loads the full detailed text from the archive when needed.

Because of this design, the agent always has the right “specialized tool” for each part of the database. It can jump quickly between titles, legal sections and technical comparisons to build the best answer.

Honestly, it still feels crazy to see what becomes possible once you start experimenting and rebuilding these things yourself.Without trying it hands on, I would have never imagined ideas like this.

And I love how clearly you can influence the quality of RAG once you understand the structure behind it.

0 comments