Building Production RAG for 10K+ NASA Docs (Scales to 85K+): Rocket Schematics and more! (Day 8)

Описание: Hey guys, I'm Raj. I've been building AI agents and RAG systems for enterprises — banks, pharma companies, legal firms, and more. Currently working on my startup (intraplex.ai).

This is what we are building: https://airstrip-ai-secure-com.notion...

In this live stream, I wanted to show how real production systems are built — or at least how I would approach it and go about building something at this scale. Building with 10,000 NASA documents, but designing the architecture to handle 85,000+ documents and beyond. Rocket schematics, technical diagrams, legacy scans from the 1970s, aerospace terminology — the whole mess.

The goal isn't just to make it work on 10K docs. It's to build the foundation that actually scales — proper chunking, metadata architecture, hybrid retrieval, vision processing for technical diagrams — so when you throw 85K or 100K+ documents at it, it doesn't fall apart.

I want to keep it raw and natural. I want to show you how much figuring out and thinking it actually requires to build something like this. The debugging, the failed approaches, the architecture decisions that matter at scale.

I'm in no way claiming to have the perfect approach or the best approach — but I wanted to share how I would go about this in a couple of days/streams. Hopefully this is helpful for some folks trying to build real systems.

Will share the GitHub once it's done.