Project Golem Fork Adds Milvus Backing

This is the rare demo that gets more valuable when it admits its own limits. The original project is a great “see why retrieval failed” toy; the Milvus fork is what it looks like when that idea graduates into something a team could actually debug against real data.

• –Turning embeddings into a 3D map makes chunking, clustering, and bad retrieval immediately legible instead of hidden behind scores.
• –The original brute-force, full-rebuild pipeline is fine for a proof of concept, but it falls apart as document counts and update frequency rise.
• –Milvus adds ANN search, metadata filters, and streaming ingestion, which is exactly what a production RAG backend needs.
• –The real product decision is tradeoff management: keep the core forkable and playful, or add features that make it less hackable but far more capable.
• –The separation of concerns is the smartest part of the fork: visualization stays lightweight while retrieval becomes an actual database problem.