Single tool 'askFabricDocs' has a clearly named verb-starting identifier and a reasonable description, but exhibits significant gaps in parameter documentation and output schema clarity. The tool description is adequate (95 chars), but parameter descriptions exist only in schema definitions (not visible in the actual handler registration context shown). The output schema is not explicitly documented, the handler returns TextContent[] but no structured metadata about what fields LLMs should expect in enriched responses. The implementation shows good error handling with actionable messages ('Embedding model not ready'), but the tool interface itself lacks guidance on output pagination, field structure, and chaining identifiers needed for downstream operations.

• Document the output schema explicitly: 'Returns an array of objects, each with: { text: string (markdown-formatted chunk), source: { filePath: string, heading: string } }. Maximum 20 results per call. For larger result sets, refine your query with more specific keywords.'
• Update parameter descriptions: change 'resultCount' description to 'The maximum number of relevant document chunks to return (1 - 20, default 8). Larger counts may include less relevant results; for best precision, use 5 - 8.'
• Add to tool description: 'Prerequisite: Fabric UX documentation must be indexed via the index-docs npm script. If no results are returned, verify indexing completed successfully by checking server logs.'
• Enhance error handling for ChromaDB query failures: catch and return actionable errors like 'No documents indexed yet. Run: npm run index-docs' or 'Collection not found in database. Indexing may have failed, check server startup logs.'
• Return structured metadata alongside documents: instead of concatenating into text, return { documents: [ { text, source: { filePath, heading } } ], totalCount, hasMore } so downstream tools can filter, sort, or link back to sources.
• Add a note to the tool description about typical latency: 'Typical response time: 500ms - 2s depending on query complexity and result count. This tool is read-only and safe to call repeatedly.'

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry