Ragdrift Mcp
Interpret RAG quality scores, tune thresholds, and understand drift dimensions while you ship retrieval features.
Overview
io.github.MukundaKatta/ragdrift-mcp is a Build-phase MCP server that diagnoses RAG drift by interpreting scores, recommending thresholds, and explaining quality dimensions.
What is this MCP server?
- Diagnoses RAG drift with interpreted scores, not raw metrics alone
- Recommends threshold adjustments for retrieval quality gates
- Explains evaluation dimensions so solo builders know what moved
- Stdio MCP @mukundakatta/ragdrift-mcp (v0.1.1) from mcp-stack
- Supports agent-led tuning loops before and after launch
What problem does it solve?
Your RAG assistant quality slipped but you only see opaque similarity numbers and do not know whether embeddings, chunks, or filters caused the drift.
Who is it for?
Indie builders running small RAG products who need structured interpretation of eval scores inside MCP.
Skip if: Teams without any RAG eval metrics or those wanting a full hosted observability product out of the box.
What do I get? / Deliverables
You get agent-readable drift diagnosis with threshold guidance so you can fix retrieval settings with a clear next experiment.
- Interpreted drift diagnosis with dimension explanations
- Threshold recommendations for retrieval quality gates
- Agent-actionable tuning notes for the next embedding or chunk experiment
Recommended MCP Servers
Journey fit
RAG setup is core build work; diagnosing drift early prevents shipping a chat feature that silently answers from stale chunks. Agent-tooling covers retrieval pipelines where the agent needs structured guidance on scores and thresholds, not just raw floats.
How it compares
RAG drift advisory MCP, not a vector DB connector or generic log monitor.
Common Questions / FAQ
Who is ragdrift-mcp for?
Solo developers shipping retrieval-augmented agents who already produce drift or quality scores and want MCP-native interpretation.
When should I use ragdrift-mcp?
Use it when eval scores shift after doc updates, model changes, or chunk policy tweaks and you need threshold and dimension guidance.
How do I add ragdrift-mcp to my agent?
Install @mukundakatta/ragdrift-mcp via npm, register the stdio MCP server, and pass your RAG evaluation inputs to the diagnose tools from your agent.