MCP server for Milvus vector database operations
The server provides 8 tools with reasonable naming and basic descriptions. However, there are significant gaps in schema completeness, parameter documentation, output schema specification, and error handling guidance. Most tools lack detailed constraint documentation and output structures are not formally documented. Parameter descriptions exist but are often minimal (10-30 chars), falling short of the 72-char production baseline. The codebase shows proper tool registration via FastMCP with JSON Schema inputs, but output schemas are not declared and error responses lack recovery guidance.
Create a new collection with quick setup or customized schema.
Get detailed information about a collection.
Perform hybrid search combining BM25 text search and vector search with RRF ranking.
List all collections in the database.
Query collection using filter expressions.
Perform full text search on a collection.
No output schemas documented for any tool. LLMs cannot predict what fields will be returned, forcing them to either guess or fail at downstream chaining. create_collection, vector_search, hybrid_search return complex nested structures but no formal schema is provided.
Parameter descriptions are often too brief (10-30 chars) and lack actionable detail. E.g. 'drop_ratio' in search_collection states 'Proportion of low-frequency terms to ignore (0.0-1.0)' but does not explain when or why to adjust it, or what happens at the extremes (0.0 vs 1.0). Production baseline is 72 chars average.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | B | 70 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
Perform text similarity search on a collection.
Perform vector similarity search on a collection.
create_collection has a complex 'field_schema' parameter accepting arbitrary JSON objects with no constraint documentation. LLMs cannot know what fields are required/optional in those objects, or how they map to Milvus DataType enums. This invites malformed payloads.
No error handling guidance in tool definitions or implementation. When search_collection, vector_search, or query_collection fail (e.g. invalid metric_type, collection not found), the LLM receives only a generic ValueError. No recovery suggestions or categorization (retryable vs user-fixable vs fatal).
create_collection is a destructive WRITE operation but has no confirmation or dry-run capability. An agent could accidentally create a collection with wrong schema and no undo path is offered.
vector_search, text_similarity_search, and hybrid_search have optional 'radius' and 'range_filter' parameters that are mutually exclusive or context-dependent, but this constraint is not documented in parameter descriptions. LLMs may pass both, causing ambiguity.
metric_type parameter in vector_search, text_similarity_search, and hybrid_search accepts 'COSINE', 'L2', 'IP' but is not formally constrained as an enum in the schema. Description lists allowed values as text, but LLMs may still hallucinate invalid options like 'HAMMING' or 'EUCLIDEAN'.
search_collection and query_collection accept output_fields arrays but no documentation explains what fields are available on a given collection or what happens if a field does not exist. LLMs cannot discover valid field names without a separate list_fields tool.
No pagination support in search_collection, query_collection, vector_search, text_similarity_search, or hybrid_search. The 'limit' parameter controls max results but no offset/page/cursor support exists. If a user query matches 10,000 vectors, the agent cannot fetch results beyond the limit.