A recommendation system MCP server built with Go Kratos framework, providing tools for product recommendation across recall, ranking, and filtering stages
Two tools with reasonable naming and descriptions, but schemas lack proper type annotations, parameter descriptions are incomplete, and there is no documented output schema. The tools follow verb_noun naming (TriggerRecommend, GetRecommendHistory), and descriptions are adequate (151 and 135 chars respectively), both within the 10 - 1024 character guideline. However, input schemas show significant gaps: nested object properties lack descriptions (e.g., 'properties' under user_feature and action_list items have no 'description' fields), and the output schema is completely undocumented. Parameter constraints are present (enums for action_type and scene), but many fields are marked 'optional' without clear defaults or consequences. Error handling and recovery guidance are absent from both tool descriptions.
Retrieves the historical recommendation results for a specific user, including recommended items and their recommendation scores
Triggers the recommendation pipeline to generate personalized product recommendations based on user context, behavior history, and scene type
Output schema completely undocumented. Neither tool declares what fields the response contains, forcing LLMs to infer structure and limiting their ability to chain calls or extract required data.
Nested object properties lack descriptions. user_feature.properties and action_list items do not have 'description' fields for user_id, age, tags, device_type, item_id, action_type, timestamp. LLMs cannot infer meaning from property names alone.
No error handling or recovery guidance in tool descriptions. Missing guidance on what to do if recommendation pipeline fails, user not found, invalid scene, or when history is empty.
Optional parameters lack default values or consequences. user_feature (except user_id), action_list, and scene_context are marked optional with no indication of how the tools behave when omitted.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | D | 56 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 31 | - | v1 |
Parameter descriptions are generic. 'User behavior action history' does not explain the expected order (chronological?), size limits, or whether duplicates are allowed. 'Additional scene context information' is too vague.
Enum values lack descriptive labels. action_type uses 0 - 4 with brief English names in the description, but the schema does not map enum integers to labels directly. The description says '0=unspecified, 1=click, 2=view, 3=collect, 4=buy', this forces LLMs to remember the mapping.
GetRecommendHistory has no constraints on limit and offset. No documented min/max for pagination parameters, risking runaway requests for limit=1000000 or negative offsets.