Hey Product Hunt 👋 I'm Max. I started Cymatix after trying to run a 20B+ parameter model on 12GB of VRAM. Once the model had eaten my memory, there was no room left for retrieval. So I built a retrieval engine that runs on the CPU alone and works with a small local model or a frontier one. What makes it different: Deterministic, math-only retrieval. There's no LLM in the hot path, so the same query gets the same result every time and you can audit it. Runs locally. Your data stays on your machine, which matters for air-gapped and offline setups. Plugs into agents through MCP, with deferred tool loading so it doesn't bloat your context. Verified 33.93% on the EnterpriseRAG Benchmark leaderboard, and collaborators have reproduced it independently on ARM64. It's open source under Apache 2.0. I'd love feedback on where retrieval breaks for you, whether that's scale, messy docs, or domain-specific corpora. That's the stuff I'm actively working on.

Forum Threads

p/cymatix-context

•

1d ago

Where would you run retrieval if it had to stay completely offline?

Cymatix Context runs retrieval on CPU with plain math (IDF-weighted, deterministic), with no model in the query path. Collaborators have already reproduced it on ARM64 hardware, which got me thinking about places where cloud RAG just isn't allowed or isn't practical: air-gapped plants, shop floors, field laptops, regulated offices, a Pi in a closet.

If you've got a box or an environment like that, what would you want to search there? And what would make or break it for you: RAM, index size, ingest speed, recall latency or something else?

View all

Cymatix Context

Same query, same answer: deterministic RAG for agents

Same query, same answer: deterministic RAG for agents

Forum Threads

Where would you run retrieval if it had to stay completely offline?

Forum Threads

Where would you run retrieval if it had to stay completely offline?