MCP Server for Semarchy xDM (Data Management) that provides resources and tools for listing data locations, retrieving model descriptions, querying data using SemQL, and generating prompts to convert natural language queries into SemQL queries.
This server has two tools with basic functional definitions but significant quality gaps. Schemas are present and typed, but descriptions are minimal and lack actionability. Parameter documentation is sparse, error handling is absent, and the tool composition creates unnecessary friction for agents. The naming is acceptable but generic. Based on the rubric baselines (avg tool description 194 chars, p10=34, p90=392), both descriptions fall in the p10 range, indicating below-average quality. No tool has comprehensive descriptions explaining when to use it vs. similar tools, what errors can occur, or how to chain outputs.
Get a model description for a given data location.
Query data from a given data location in natural language.
Tool descriptions are below rubric baseline. 'Get a model description for a given data location.' (50 chars) lacks WHEN to use it, WHY it's needed before query_data, and WHAT format is returned. Rubric requires 10-1024 chars with WHAT/WHEN/prerequisites. This is p10 (minimal) quality.
query_data description (67 chars) is vague: 'Query data from a given data location in natural language.' Does not explain that a SemQL query OBJECT (not natural language) is required, does not describe what 'entity' is mandatory, and does not document what fields the response contains. Agents cannot plan from this description alone.
Parameter descriptions in query_data are inconsistent. The 'query' parameter has a description ('A SemQL query object to execute') but the nested Pydantic model SemqlQuery has per-field descriptions in code (DMClient.py), NOT visible in the tool schema. LLMs see the outer description only; they cannot see that 'entity' is mandatory or that 'expressions' can be None.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 48 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 45 | - | v1 |
No output schema documented. Both tools return strings (model description, query results). The rubric requires documentation of return type and structure (fields the agent can extract). Without output schemas, agents cannot plan downstream actions or extract relevant data from responses.
No error handling guidance. Both tools can fail (model not found, query returns no data, API errors). Responses include fallback strings ('Model for data location {datalocation} not found.') but no guidance on what the LLM should do next (retry? Search for available models? Ask the user?). Rubric requires 'recovery guide' pattern.
Tool composition forces multi-step discovery. The prompt 'generate_query_prompt' instructs the LLM to first call get_model_description, then call query_data. This should be implicit in the tool signatures. Additionally, get_model_description modifies global state (dm_client.default_data_location = datalocation), which is non-idempotent and breaks agent retry logic.
query_data parameter 'query' is typed as SemqlQuery (Pydantic model), but the tool expects a complex object with nested optional fields. The description does not clarify which fields are required (entity is mandatory per code) vs optional. Agents cannot reliably construct valid SemQL queries from this definition.
get_model_description takes 'datalocation' (string) but does not describe valid values, format, or how to discover available data locations. Users/agents see a resource 'semarchy://data-locations' that returns a formatted string, not a structured list. This forces manual parsing or extra discovery friction.