Review: Ratel, context engineering for production agents

These are vendor-run benchmarks on a synthetic catalog, and that last figure is the one to read carefully: 8.3% is roughly what you'd expect when a model is drowning in undifferentiated tool schemas, so the delta is as much a statement about how bad the naive baseline is as about how good the retrieval is. The direction is almost certainly real — retrieval over a large tool catalog beats stuffing it — but treat the magnitudes as an upper bound until you've run your own catalog through it.

#What's good

The zero-setup default is the right call. BM25 in-process means you can put this in front of an existing agent without provisioning anything, and only reach for embeddings if keyword search underperforms on your catalog. Most context-engineering products get this backwards and lead with a vector store.

Traces are a first-class feature, not an afterthought. "Why did the agent choose that tool" is the single hardest question to answer when a production agent misbehaves, and shipping OpenTelemetry spans for retrieval decisions is the difference between debugging and guessing.

Open-source engine. Apache-2.0 on the part that does the work means the lock-in question has a real answer.

#What's rough

• Memory is marketed ahead of the docs. The site pitches unified memory and cross-agent "fleet learning"; the documentation lists agent memory — and chat-history management — as coming soon. If memory is why you're evaluating it, verify what ships today before building on it.
• Self-improving ranking is experimental, which means ranking behaviour can drift between runs. That's a hard property to reason about in production; you'll want the traces on.
• Public pricing is absent. The commercial side is a "contact us" motion, so budget for a conversation.
• It adds a hop. Retrieval before every turn is another component in the request path and another thing to keep warm — worth it above a certain catalog size, pure overhead below it.

#Verdict

If your agent has ten tools, skip it — you don't have the problem this solves. If it has a hundred, spans several MCP servers, or you're paying real money for tokens spent re-sending schemas the model never calls, this is the most credible open-source attempt at the problem right now, and the open engine plus BM25 default makes the trial cheap. Run their benchmark harness against your catalog before you believe any specific percentage.

Sources:

• Ratel — ratel.sh
• Ratel documentation