Query precomputed PPE compliance analytics from a factory safety video.
Server has 11 well-named tools with consistent verb_noun patterns (get_, list_, explain_). All tools have descriptions (avg ~80 chars, within baseline 34-392). Input schemas are present and typed for all tools. However, descriptions are terse and lack context on WHEN to use each tool or what downstream actions they enable. Parameter descriptions are minimal (e.g., 'Worker track ID' for track_id). Output schemas are not documented, LLMs cannot predict response structure for chaining. Error handling returns basic dicts with 'error' fields but lacks recovery guidance. No tool annotations (readOnlyHint, destructiveHint). Resource endpoints exist but are underdocumented.
Human-readable compliance narrative plus streak stats for one worker.
Site compliance rate over time buckets (default 2 second windows).
Short-lived tracker IDs that may be noise (seen in at most max_frames).
Site-wide PPE compliance summary for the analyzed video.
Frame-by-frame compliance events, optionally filtered by worker or frame range.
Workers with PPE violations between start_s and end_s (seconds).
Final compliance status and presence stats for one tracked worker ID.
Output schemas not documented. LLMs cannot predict response structure (fields, types, nesting) for downstream chaining. E.g., get_worker returns a dict but no schema tells the LLM what keys to expect.
Descriptions lack WHEN-to-use context and recovery guidance. E.g., 'Final compliance status and presence stats for one tracked worker ID' does not explain when to call get_worker vs explain_worker, or what to do if track_id is invalid.
Error handling returns bare error dicts without recovery guidance. E.g., KeyError for missing track_id returns {'error': '...', 'track_id': ...} but does not suggest calling list_violations or get_worker_summary_ranked to find valid IDs.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
All workers ranked by helmet_ratio, vest_ratio, time_on_site_s, etc.
Top N workers with the worst PPE violations by severity score.
List workers with PPE violations. Optionally filter by helmet or safety-vest.
Reload analytics JSON from disk after a new inference run.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). All tools are read-only but this is not declared, forcing LLMs to infer safety from descriptions.
Parameter descriptions are minimal. E.g., 'Worker track ID' for track_id does not explain valid range, how to discover valid IDs, or what happens if ID is invalid.