Most RAG failures don’t crash. They silently return bad answers. I built a repair layer for that.

Most RAG tooling provides a score but fails to specify what actually went wrong.

I had retrieval failures, grounding issues, generation going sideways, all showing up as a number. No way to know which failure caused which run to go wrong. No way to fix it without guessing.

So I built ragbolt.

ragbolt is a failure-aware repair layer for RAG pipelines that:

• Detects whether the failure originated from retrieval, generation, or grounding
• Applies one bounded repair at a time
• Re-verifies the result
• Emits a full trace to show exactly what changed and why

It’s not a framework.
Not an agent.
Not "self-healing RAG".

Just a small wrapper around existing RAG pipelines with explicit repair limits, auditability, and a hard stop when confidence breaks down.

It runs standalone and integrates with LangChain + LlamaIndex.

pip install ragbolt