A walkthrough of building documentation MCP servers for IBM Bob (an AI coding assistant), using the watsonx Orchestrate documentation server as a reference. The post covers designing two tools: a hybrid semantic+keyword search tool that returns concise section-level metadata, and a virtual filesystem query_docs tool that lets agents read specific line ranges from markdown files using shell-like commands. LanceDB is used as the lightweight embedded search engine. The design goal is to minimize context window usage by returning only relevant document sections rather than full pages.
99 Impressions