Model Context Protocol server for CodeSageZ Graph-Augmented RAG. Exposes graph-augmented RAG tools for integration with Claude Code, Cursor, and MCP clients.
Single tool with adequate naming and schema structure, but significant gaps in parameter descriptions and output documentation. Tool name 'retrieve_code_context' follows verb_noun convention and clearly conveys action. Input schema present with three parameters (repo_id, query, mode), all typed. However, parameter descriptions are minimal (10-30 chars each), falling below the 72-char baseline for production tools. Output schema (RetrieveResponse) is defined but lacks field-level descriptions. Tool description (95 chars) is below the 194-char average for production tools and omits WHEN to use this tool vs alternatives. No error handling guidance, no recovery paths, no validation rules documented. HTTP transport is current, but tool definition lacks LLM-optimization patterns.
Retrieves relevant code chunks using Graph-Augmented RAG (1-hop topological call-graph expansion). Graph expansion surfaces callee/caller context that pure vector search misses.
Parameter descriptions are too brief (10-30 chars vs 72-char baseline). 'repo_id' described as 'Ingested repository UUID or name' lacks format guidance, valid value examples, or constraints. 'query' and 'mode' similarly underdescribed.
Output schema (RetrieveResponse) lacks field-level descriptions. 'chunks' is a list of dicts with no documented structure. 'latency_ms', 'mode', 'recall_estimate' have no descriptions. LLMs cannot infer what fields to extract or how to use them downstream.
Tool description (95 chars) is below production baseline (194 chars) and lacks WHEN/WHY guidance. Does not explain when to call this vs a naive vector search, what prerequisites exist (repo must be ingested?), or what 'graph' mode returns vs 'naive'.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
No error handling or recovery guidance. If repo_id is invalid, query is empty, or mode is unrecognized, no documented error response or next steps. LLM has no guidance on retry vs user correction.
'mode' parameter has a default ('graph') but no enum constraint. LLM could pass invalid modes like 'hybrid' or 'semantic'. Should declare mode as enum: ['graph', 'naive'].