Infrastructure Q&A agent over MCP

Hallucinations traced to the tool layer · tool catalog in context cut in half

What was built

An agent you ask about infrastructure in plain words. LLM with tool-calling over MCP servers for inventory and monitoring, answers in a bot. I ran 15 test questions, checking each answer against the database. Six held up. Not the model: narrow query schema, no honest counting, step limits choking it. All tool layer. That's where it got fixed. The catalog is resent every iteration, so I halved it. In production. No demand yet.

Stack

MCPLLM tool-callingPythonTelegram Bot APIDockerGitLab CI

Other work