An MCP server for GraphRAG that provides tools for querying a Neo4j knowledge graph, retrieving unit information, finding similar units by image, and accessing ontology data.
Critical gaps across all dimensions. Tool descriptions are present but mostly in German and lack LLM-optimized structure (10-50 chars, below 194-char baseline). Input schemas exist but are minimal: most tools have 0-1 parameters with no type constraints or enums. Parameter descriptions are sparse or missing entirely. No output schemas documented. No error handling guidance. Tool naming is reasonable (verb_noun pattern) but composition issues exist: get_similar_units_by_asset_id, get_similar_units_by_image, and get_similar_units_by_image_url do nearly identical things, forcing LLM disambiguation. No pagination, no result limits stated. Security: no audit logging visible, no permission gates, no rate limiting.
Sucht anhand der asset_id des zuvor hochgeladenen Bilds nach den maximal 3 ähnlichsten Bildern im Knowledge Graph.
Sucht anhand des zuvor hochgeladenen Bilds nach den maximal 3 ähnlichsten Bildern im Knowledge Graph.
Ruft den tiefen epistemischen Kontext für ein zuvor auf dem Server bereitgestelltes Bild ab. Erwartet die eindeutige Asset-ID (UUID).
Ruft den tiefen Kontext für eine oder mehrere Units über den MCP-Server ab. Akzeptiert einzelne Unit-IDs (Int) oder Listen davon.
Provides onboarding briefing prompt for the GraphRAG system
Führt eine Cypher-Query in der Neo4j-Datenbank aus. Beachte, dass der Graph read-only ist.
Three nearly-identical similarity search tools (get_similar_units_by_asset_id, get_similar_units_by_image, get_similar_units_by_image_url) force LLM disambiguation and violate single-responsibility principle. Should consolidate into one tool with overloaded parameters or clearly document when each is appropriate.
get_similar_units_by_image has empty input schema (no required parameters), but description says 'zuvor hochgeladenen Bilds' (previously uploaded image). Unclear how the tool accesses the image, is it session state? Stateful behavior violates MCP statelessness principle.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields. Baseline: 100% of A+ tools have documented return types.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 46 | 2026-07-28+ | v2 |
Descriptions are in German and lack LLM-optimized structure. 'Führt eine Cypher-Query...' (35 chars) is below 194-char baseline. No WHAT/WHEN/WHY structure. No dependency hints (e.g., 'Call read_graph first to explore schema'). No error recovery guidance.
No error handling guidance. Tools return READ_ONLY risk classification but no documentation of what errors are possible, how to recover, or what the LLM should do if a query fails.