ragX
Our retrieval engine: answers from our own documents with citations, and a verifier that makes it abstain when the sources do not support a claim.
The problem
A retrieval system that always answers will sometimes answer from nothing. For internal knowledge, a confident wrong answer costs more than no answer.
What we engineered
Make abstaining a first-class result. Retrieve broadly, rerank, then check every claim against the retrieved text before it is shown.
- Hybrid retrieval: dense vectors and keyword search fused by reciprocal rank fusion, then a reranker.
- A critic pass on a local model, and a natural-language-inference verifier that must entail each claim or the answer abstains with a reason.
- Indexing and evaluation jobs for the collections our tools use.
How it connected
- One service with three front doors: an HTTP API, a command line and an MCP server.
- Our coding agents query it through a prompt hook before they answer questions about our own systems.
What technology was appropriate
Python; runs locally on our own hardware with a local model for the critic step.
Where it stands
Internal and in daily use. Not offered as a hosted product; the same architecture is what we build for clients who need answers from their own documents.
Current state, stated as it is today.
See it yourself
Automate operations
Workflow automation, AI agents and knowledge assistants that take repeated work off your team.
Knowledge systems and RAG
Answers grounded in your own documents, with citations and a refusal when the source is missing.
Private, local and cloud AI
Run AI on your devices, your servers, a private cloud or the public cloud, chosen per project.
Bring us the bottleneck.
If your problem looks like one of these, tell us. We will say whether a build is the right answer.
Bring us a bottleneck