Start to learn vocab from talkgroups to improve accuracy of STT

drb-c2-core/app/internal/intelligence.py

@@ -36,7 +36,7 @@ Rules:
 
 System: {system_id}
 Talkgroup: {talkgroup_name}
-{transcript_block}"""
+{vocabulary_block}{transcript_block}"""
 
 # Nominatim viewbox half-width in degrees (~11 km at mid-latitudes)
 _GEO_DELTA = 0.1
@@ -76,8 +76,15 @@ async def extract_tags(
     Side-effect: updates calls/{call_id} in Firestore with tags, location,
     location_coords, vehicles, units, severity, transcript_corrected; also stores embedding.
     """
+    # Load per-system vocabulary for prompt injection
+    vocabulary: list[str] = []
+    if system_id:
+        from app.internal.vocabulary_learner import get_vocabulary
+        vocab_data = await get_vocabulary(system_id)
+        vocabulary = vocab_data.get("vocabulary") or []
+
     result = await asyncio.to_thread(
-        _sync_extract, transcript, talkgroup_name, talkgroup_id, system_id, segments
+        _sync_extract, transcript, talkgroup_name, talkgroup_id, system_id, segments, vocabulary
     )
 
     tags: list[str]              = result.get("tags") or []
@@ -233,6 +240,7 @@ def _sync_extract(
     talkgroup_id: Optional[int],
     system_id: Optional[str],
     segments: Optional[list[dict]],
+    vocabulary: Optional[list[str]] = None,
 ) -> dict:
     """Call GPT-4o mini and parse the JSON response."""
     from app.config import settings
@@ -242,11 +250,13 @@ def _sync_extract(
         logger.warning("OPENAI_API_KEY not set — intelligence extraction disabled.")
         return {}
 
+    from app.internal.vocabulary_learner import build_gpt_vocab_block
     tg = f"{talkgroup_name} (TGID {talkgroup_id})" if talkgroup_id else (talkgroup_name or "unknown")
     prompt = _PROMPT_TEMPLATE.format(
         transcript_block=_build_transcript_block(transcript, segments),
         talkgroup_name=tg,
         system_id=system_id or "unknown",
+        vocabulary_block=build_gpt_vocab_block(vocabulary or []),
     )
 
     try: