Fix GPT thread safety, improve relevance logic, and add caching for COCO matching

app.py

@@ -90,16 +90,15 @@ async def _enrich_first_frame_gpt(
         # LLM relevance filter (if LLM_EXTRACTED mode)
         gpt_dets = detections
         if mission_spec and mission_spec.parse_mode == "LLM_EXTRACTED":
-            if not mission_spec.relevance_criteria.required_classes:
-                unique_labels = list({
-                    d.get("label", "").lower()
-                    for d in detections if d.get("label")
-                })
-                relevant_labels = await asyncio.to_thread(
-                    evaluate_relevance_llm, unique_labels, mission_spec.operator_text
-                )
-                mission_spec.relevance_criteria.required_classes = list(relevant_labels)
-            # Apply deterministic filter
+            unique_labels = list({
+                d.get("label", "").lower()
+                for d in detections if d.get("label")
+            })
+            relevant_labels = await asyncio.to_thread(
+                evaluate_relevance_llm, unique_labels, mission_spec.operator_text
+            )
+            mission_spec.relevance_criteria.required_classes = list(relevant_labels)
+            # Apply deterministic filter with refined classes
             for d in detections:
                 decision = evaluate_relevance(d, mission_spec.relevance_criteria)
                 d["mission_relevant"] = decision.relevant

coco_classes.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import difflib
+import functools
 import logging
 import re
 from typing import Dict, Optional, Tuple
@@ -219,6 +220,7 @@ def _semantic_coco_match(value: str) -> Optional[str]:
     return None
 
 
+@functools.lru_cache(maxsize=512)
 def canonicalize_coco_name(value: str | None) -> str | None:
     """Map an arbitrary string to the closest COCO class name if possible.
 

inference.py

@@ -1046,6 +1046,7 @@ def run_inference(
     # --- GPT Enrichment Thread (non-blocking) ---
     # Runs LLM relevance + GPT threat assessment off the writer's critical path.
     gpt_enrichment_queue = Queue(maxsize=4)
+    _relevance_refined = [False]  # mutable container for thread-safe sharing
 
     def enrichment_thread_fn(tracker_ref):
         """Dedicated thread for GPT/LLM calls. Receives work from writer, injects results via tracker."""
@@ -1055,9 +1056,8 @@ def run_inference(
                 break  # Sentinel — shutdown
             frame_idx, frame_data, gpt_dets, ms = item
             try:
-                # LLM post-filter (LLM_EXTRACTED mode, frame 0 only)
-                if (ms and ms.parse_mode == "LLM_EXTRACTED"
-                        and not ms.relevance_criteria.required_classes):
+                # LLM post-filter (LLM_EXTRACTED mode)
+                if ms and ms.parse_mode == "LLM_EXTRACTED":
                     unique_labels = list({
                         d.get("label", "").lower()
                         for d in gpt_dets if d.get("label")
@@ -1066,10 +1066,16 @@ def run_inference(
                         unique_labels, ms.operator_text
                     )
                     ms.relevance_criteria.required_classes = list(relevant_labels)
+                    _relevance_refined[0] = True  # signal writer_loop to switch to deterministic gate
                     logging.info(
                         "Enrichment: LLM post-filter applied on frame %d: relevant=%s",
                         frame_idx, relevant_labels,
                     )
+                    # Re-filter with refined classes
+                    for d in gpt_dets:
+                        decision = evaluate_relevance(d, ms.relevance_criteria)
+                        d["mission_relevant"] = decision.relevant
+                    gpt_dets = [d for d in gpt_dets if d.get("mission_relevant", True)]
 
                 # GPT threat assessment
                 if gpt_dets:
@@ -1151,30 +1157,39 @@ def run_inference(
 
                         # --- RELEVANCE GATE (deterministic, fast — stays in writer) ---
                         if mission_spec:
-                            for d in dets:
-                                decision = evaluate_relevance(d, mission_spec.relevance_criteria)
-                                d["mission_relevant"] = decision.relevant
-                                d["relevance_reason"] = decision.reason
-                                if not decision.relevant:
-                                    logging.info(
-                                        json_module.dumps({
-                                            "event": "relevance_decision",
-                                            "track_id": d.get("track_id"),
-                                            "label": d.get("label"),
-                                            "relevant": False,
-                                            "reason": decision.reason,
-                                            "required_classes": mission_spec.relevance_criteria.required_classes,
-                                            "frame": next_idx,
-                                        })
-                                    )
-                            gpt_dets = [d for d in dets if d.get("mission_relevant", True)]
+                            if (mission_spec.parse_mode == "LLM_EXTRACTED"
+                                    and not _relevance_refined[0]):
+                                # LLM post-filter hasn't run yet — pass all through
+                                for d in dets:
+                                    d["mission_relevant"] = True
+                                    d["relevance_reason"] = "pending_llm_postfilter"
+                                gpt_dets = dets
+                            else:
+                                # Normal deterministic gate (with refined or FAST_PATH classes)
+                                for d in dets:
+                                    decision = evaluate_relevance(d, mission_spec.relevance_criteria)
+                                    d["mission_relevant"] = decision.relevant
+                                    d["relevance_reason"] = decision.reason
+                                    if not decision.relevant:
+                                        logging.info(
+                                            json_module.dumps({
+                                                "event": "relevance_decision",
+                                                "track_id": d.get("track_id"),
+                                                "label": d.get("label"),
+                                                "relevant": False,
+                                                "reason": decision.reason,
+                                                "required_classes": mission_spec.relevance_criteria.required_classes,
+                                                "frame": next_idx,
+                                            })
+                                        )
+                                gpt_dets = [d for d in dets if d.get("mission_relevant", True)]
                         else:
                             for d in dets:
                                 d["mission_relevant"] = None
                             gpt_dets = dets
 
                         # --- GPT ENRICHMENT (non-blocking, offloaded to enrichment thread) ---
-                        if next_idx == 0 and enable_gpt and gpt_dets and not gpt_submitted:
+                        if enable_gpt and gpt_dets and not gpt_submitted:
                             # Tag as pending — enrichment thread will update to ASSESSED later
                             for d in gpt_dets:
                                 d["assessment_status"] = "PENDING_GPT"
@@ -1184,7 +1199,7 @@ def run_inference(
                                     timeout=1.0,
                                 )
                                 gpt_submitted = True
-                                logging.info("Writer: offloaded GPT enrichment for frame 0")
+                                logging.info("Writer: offloaded GPT enrichment for frame %d", next_idx)
                             except Full:
                                 logging.warning("GPT enrichment queue full, skipping frame 0 GPT")