realroc/shumei-user-violation-audit

Shumei User Violation Audit

Use this skill to turn a user cohort plus conversation-selection rule into a per-user CSV: 用户 -> conversation.query -> Shumei input event -> 违禁请求数 / 检测成功请求数.

Required Inputs

Before running, identify these values from the user or local context:

• User filter: Mongo query for the user collection.
• Conversation filter: Mongo query template for the conversation collection.
• Shumei key: accessKey for the text risk service.
• Mongo access: either direct mongo_uri + mongo_db, or an app/container environment that already has Mongo access.
• Output path: default to the current workspace with a timestamped CSV when not specified.

Do not hardcode the Shumei key in the skill. Treat it as run input.

Script

Use the bundled script:

python3 scripts/audit_user_violation_rate.py --config config.json

Print an editable config template:

python3 scripts/audit_user_violation_rate.py --print-example-config

The script is read-only for MongoDB. It writes a UTF-8-BOM CSV and a .gz copy unless --no-gzip is passed.

Filter Templates

The script accepts valid JSON. Template placeholders are resolved before querying:

• {{now}}
• {{now - 2592000}}
• {{username}}
• {{register_time}}
• {{register_time + 86400}}
• {{user.some_nested_field}}

Default conversation template matches the prior AMA audit:

{
  "user": "{{username}}",
  "request_time": {
    "$gte": "{{register_time}}",
    "$lt": "{{register_time + 86400}}"
  },
  "query": { "$exists": true, "$ne": "" }
}

Example still-free users registered in the past 30 days:

{
  "register_time": { "$gte": "{{now - 2592000}}", "$lt": "{{now}}" },
  "$and": [
    {
      "$or": [
        { "pay_history": { "$exists": false } },
        { "pay_history": [] },
        { "pay_history": null }
      ]
    },
    {
      "$or": [
        { "vip_expire_time": { "$exists": false } },
        { "vip_expire_time": { "$lte": "{{now}}" } },
        { "vip_expire_time": -1 }
      ]
    }
  ]
}

Workflow

1. Convert the user's cohort requirement into user_filter.
2. Convert the conversation requirement into conversation_filter_template; include query: {"$exists": true, "$ne": ""} unless the user explicitly wants empty queries included.
3. Create a temporary config JSON outside the skill directory, usually in the workspace.
4. Run the script where Mongo is reachable.
5. Verify the final RESULT_JSON_START summary: row count, request count, violation count, and failed checks.
6. Return the CSV path and summarize the key counts.

If the target MongoDB is only reachable from a server/container, copy or mount the script there and run it inside that environment. For very large CSVs over TAT/stdout, prefer leaving the report file on the server or transferring a compressed .csv.gz; avoid printing CSV contents to stdout.

Output Semantics

The default violation rule follows the existing product behavior:

code == 1100 and riskLevel != PASS

CSV rows are sorted by:

1. 请求违禁率 descending
2. 违禁请求数 descending
3. 对话请求数 descending

Default columns include user identity fields, request counts, Shumei risk labels/descriptions, up to three violating query examples, and first/last selected request times.