correlator: LLM tier reads the scene transcript, not the whole call (#112)

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

@@ -670,6 +670,7 @@ async def correlate_call(
     reassignment: bool = False,
     embedding: Optional[list] = None,
     severity: Optional[str] = None,
+    transcript: Optional[str] = None,
 ) -> Optional[str]:
     """
     Link call_id to an existing incident or create a new one.
@@ -686,7 +687,7 @@ async def correlate_call(
         system_id=system_id, talkgroup_id=talkgroup_id, talkgroup_name=talkgroup_name,
         tags=tags, incident_type=incident_type, location=location,
         reassignment=reassignment, create_if_new=create_if_new,
-        embedding=embedding, severity=severity,
+        embedding=embedding, severity=severity, transcript=transcript,
     )
     decision = _run_decision(ctx)
     return await _apply_and_log(decision, ctx)
@@ -710,6 +711,7 @@ async def preview_correlation(
     reassignment: bool = False,
     embedding: Optional[list] = None,
     severity: Optional[str] = None,
+    transcript: Optional[str] = None,
 ) -> dict:
     """
     Run the rules engine and return the decision WITHOUT committing to Firestore.
@@ -730,7 +732,7 @@ async def preview_correlation(
         system_id=system_id, talkgroup_id=talkgroup_id, talkgroup_name=talkgroup_name,
         tags=tags, incident_type=incident_type, location=location,
         reassignment=reassignment, create_if_new=create_if_new,
-        embedding=embedding, severity=severity,
+        embedding=embedding, severity=severity, transcript=transcript,
     )
     decision = _run_decision(ctx)
     return {"decision": decision, "ctx": ctx}
@@ -765,6 +767,7 @@ async def _build_context(
     create_if_new: bool,
     embedding: Optional[list] = None,
     severity: Optional[str] = None,
+    transcript: Optional[str] = None,
 ) -> dict:
     now    = reference_time or datetime.now(timezone.utc)
     window = timedelta(hours=settings.correlation_window_hours)
@@ -804,6 +807,13 @@ async def _build_context(
     call_vehicles  = vehicles      if vehicles      is not None else (call_doc.get("vehicles")      or [])
     call_cleared   = cleared_units if cleared_units is not None else (call_doc.get("cleared_units") or [])
     call_severity  = severity or "routine"
+    # The transcript the LLM correlation tier reasons over. Prefer the SCENE's
+    # own words (server-26#102) — passed by upload.py's scene loop — and fall
+    # back to the call doc only when no scene text was supplied (the
+    # recorrelation sweep, and single-scene calls where the two are identical).
+    # Without this, every non-primary scene of a multi-scene call was judged by
+    # the LLM against a transcript containing the OTHER scenes.
+    scene_transcript = transcript or call_doc.get("transcript_corrected") or call_doc.get("transcript")
     # A string that is not a place is not a location anywhere downstream — not
     # in the fit tests, not in the thin-call test, not in the LLM prompt, and
     # not on the incident.  Its coordinates go with it: coords are geocoded
@@ -826,6 +836,7 @@ async def _build_context(
     return {
         "call_id": call_id, "org_id": org_id, "all_active": all_active, "recent": recent,
         "call_doc": call_doc, "call_embedding": call_embedding,
+        "scene_transcript": scene_transcript,
         "call_units": call_units, "call_vehicles": call_vehicles,
         "call_cleared": call_cleared, "call_severity": call_severity,
         "coords": coords, "is_thin_call": is_thin_call, "now": now,

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

@@ -172,7 +172,7 @@ async def extract_scenes(
 
     Each scene dict contains:
         tags, incident_type, location, location_coords, resolved,
-        severity, vehicles, units, transcript_corrected,
+        severity, vehicles, units, transcript, transcript_corrected,
         segment_indices, embedding
 
     Side-effect: updates calls/{call_id} in Firestore with merged tags,
@@ -337,6 +337,10 @@ async def extract_scenes(
         )
         embedding = await asyncio.to_thread(_sync_embed, scene_text)
 
+        scene_transcript = _scene_transcript_text(
+            transcript, segments, segment_indices, transcript_corrected
+        )
+
         processed.append({
             "tags":                 tags,
             "incident_type":        incident_type,
@@ -348,6 +352,7 @@ async def extract_scenes(
             "severity":             severity,
             "resolved":             resolved,
             "reassignment":         reassignment,
+            "transcript":           scene_transcript,
             "transcript_corrected": transcript_corrected,
             "segment_indices":      segment_indices,
             "embedding":            embedding,
@@ -571,11 +576,49 @@ def _municipality_from_tg(tg_name: Optional[str]) -> Optional[str]:
 def _build_transcript_block(transcript: str, segments: Optional[list[dict]]) -> str:
     """Format transcript as numbered transmissions if segments are available."""
     if segments and len(segments) > 1:
-        lines = [f"{i+1}. [{s['start']}s] {s['text']}" for i, s in enumerate(segments)]
+        # 0-based labels, matching the prompt's "0-based indices into the
+        # numbered transmissions" — the model echoes these back as
+        # `segment_indices`, which _build_scene_embed_text and the per-scene
+        # `transcript` (server-26#102) then slice with directly.
+        lines = [f"{i}. [{s['start']}s] {s['text']}" for i, s in enumerate(segments)]
         return f"Transmissions ({len(segments)}):\n" + "\n".join(lines)
     return f"Transcript:\n{transcript}"
 
 
+def _scene_transcript_text(
+    transcript: str,
+    segments: Optional[list[dict]],
+    segment_indices: Optional[list[int]],
+    transcript_corrected: Optional[str],
+) -> str:
+    """
+    This scene's own words, unprefixed — the segments it owns, joined.
+
+    server-26#102: the correlator's LLM tier reads this per scene instead of
+    the call doc's whole-call transcript, so on a multi-scene call scene N is
+    no longer judged against scenes 1..N-1's text.
+
+    Never returns "". Anything that would leave the slice empty — no
+    `segment_indices` (a single-segment call is never numbered by
+    `_build_transcript_block`), or indices that are out of range / not ints —
+    falls back to the whole-call transcript, which for a single-scene call is
+    the same text and for a mis-sliced multi-scene call is at least this
+    call's own words. `_sync_extract`'s prompt documents 0-based indices and
+    `_build_transcript_block` numbers to match, so no base normalisation here.
+    """
+    if transcript_corrected:
+        return transcript_corrected
+    if segments and segment_indices:
+        joined = " ".join(
+            segments[i]["text"]
+            for i in segment_indices
+            if isinstance(i, int) and 0 <= i < len(segments)
+        )
+        if joined:
+            return joined
+    return transcript
+
+
 def _build_scene_embed_text(
     transcript: str,
     segments: Optional[list[dict]],

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

@@ -61,7 +61,13 @@ def _inc_summary(inc: dict, now: datetime) -> str:
 def _call_block(ctx: dict) -> str:
     lines = []
     call_doc = ctx["call_doc"]
-    transcript = call_doc.get("transcript_corrected") or call_doc.get("transcript")
+    # The SCENE's own transcript, resolved in _build_context (server-26#102).
+    # Falls back to the call doc for a ctx built without a scene (tests, sweep).
+    transcript = (
+        ctx.get("scene_transcript")
+        or call_doc.get("transcript_corrected")
+        or call_doc.get("transcript")
+    )
     if transcript:
         lines.append(f"Transcript: {transcript[:700]}")
     if ctx["tags"]:

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

@@ -108,6 +108,7 @@ async def _recorrelate_orphan(call: dict) -> bool:
         cleared_units  = call.get("cleared_units") or [],
         embedding      = call.get("embedding"),
         severity       = call.get("severity"),
+        transcript     = call.get("transcript_corrected") or call.get("transcript"),
         reference_time = started_at,   # anchor window to when the call happened
         create_if_new  = False,        # never create — link-only
     )

drb-c2-core/app/routers/upload.py

@@ -116,6 +116,7 @@ async def _correlate_with_consensus(
     reassignment: bool = False,
     embedding: Optional[list] = None,
     severity: Optional[str] = None,
+    transcript: Optional[str] = None,
 ) -> Optional[str]:
     """
     Consensus correlator: runs the rules engine and the cheap LLM in sequence.
@@ -133,7 +134,7 @@ async def _correlate_with_consensus(
         tags=tags, incident_type=incident_type, location=location,
         location_coords=location_coords, units=units, vehicles=vehicles,
         cleared_units=cleared_units, reassignment=reassignment,
-        embedding=embedding, severity=severity,
+        embedding=embedding, severity=severity, transcript=transcript,
     )
     ctx             = preview["ctx"]
     rules_decision  = preview["decision"]
@@ -226,6 +227,7 @@ async def _run_extraction_pipeline(
                 reassignment=is_reassignment,
                 embedding=scene.get("embedding"),
                 severity=scene.get("severity"),
+                transcript=scene.get("transcript"),
             )
             if incident_id and incident_id not in incident_ids:
                 incident_ids.append(incident_id)
@@ -343,6 +345,7 @@ async def _run_intelligence_pipeline(
                 reassignment=is_reassignment,
                 embedding=scene.get("embedding"),
                 severity=scene.get("severity"),
+                transcript=scene.get("transcript"),
             )
             if incident_id and incident_id not in incident_ids:
                 incident_ids.append(incident_id)

drb-c2-core/tests/test_incident_identity.py

@@ -304,6 +304,39 @@ async def test_a_scene_is_judged_on_its_own_embedding_and_severity():
     assert ctx["call_severity"] == "major"
 
 
+@pytest.mark.asyncio
+async def test_the_llm_tier_reads_the_scene_transcript_not_the_whole_call():
+    """
+    server-26#102. intelligence.py writes only the primary scene's corrected
+    text to calls/{id}. _call_block (the LLM correlation prompt) must reason
+    over the SCENE being correlated, not a whole-call transcript that also
+    contains the other scenes. _build_context threads the scene's text in;
+    with no scene text it falls back to the call doc (sweep / single-scene).
+    """
+    with patch("app.internal.incident_correlator.fstore") as mock_fstore:
+        mock_fstore.doc_get = AsyncMock(return_value={
+            "transcript": "scene one about a fire. scene two about a traffic stop.",
+        })
+        mock_fstore.collection_list = AsyncMock(return_value=[])
+        scene = await _build_context(
+            call_id="call-1", units=None, vehicles=None, cleared_units=None,
+            location_coords=None, reference_time=NOW,
+            system_id="sys-1", talkgroup_id=383, talkgroup_name=DISPATCH_TG,
+            tags=[], incident_type="police", location=None,
+            reassignment=False, create_if_new=True,
+            transcript="scene two about a traffic stop.",
+        )
+        fallback = await _build_context(
+            call_id="call-1", units=None, vehicles=None, cleared_units=None,
+            location_coords=None, reference_time=NOW,
+            system_id="sys-1", talkgroup_id=383, talkgroup_name=DISPATCH_TG,
+            tags=[], incident_type="police", location=None,
+            reassignment=False, create_if_new=True,
+        )
+    assert scene["scene_transcript"] == "scene two about a traffic stop."
+    assert fallback["scene_transcript"] == "scene one about a fire. scene two about a traffic stop."
+
+
 @pytest.mark.asyncio
 async def test_a_bare_number_never_becomes_an_incident_location_or_title():
     inc = await _create(tags=["flames"], location="49", coords=None,

drb-c2-core/tests/test_scene_transcript.py

@@ -0,0 +1,40 @@
+"""
+server-26#102 — a scene is correlated on its OWN transcript, not the whole call.
+
+_scene_transcript_text slices the segments a scene owns. It must never return
+"" (an empty slice would let incident_correlator._build_context fall back to
+the call doc's whole-call transcript, re-opening the leak in exactly the case
+— bad indices — where it matters).
+"""
+from app.internal.intelligence import _scene_transcript_text
+
+SEGS = [
+    {"text": "structure fire, 12 Main"},
+    {"text": "engine 4 responding"},
+    {"text": "traffic stop, plate ABC"},
+    {"text": "one occupant"},
+]
+WHOLE = "structure fire, 12 Main engine 4 responding traffic stop, plate ABC one occupant"
+
+
+def test_scene_owns_a_subset_of_segments():
+    assert _scene_transcript_text(WHOLE, SEGS, [0, 1], None) == "structure fire, 12 Main engine 4 responding"
+    assert _scene_transcript_text(WHOLE, SEGS, [2, 3], None) == "traffic stop, plate ABC one occupant"
+
+
+def test_corrected_text_wins_when_present():
+    assert _scene_transcript_text(WHOLE, SEGS, [0], "cleaned up text") == "cleaned up text"
+
+
+def test_no_segment_indices_falls_back_to_whole_call():
+    # single-segment calls are never numbered by _build_transcript_block → null indices
+    assert _scene_transcript_text(WHOLE, SEGS, None, None) == WHOLE
+    assert _scene_transcript_text(WHOLE, None, [0, 1], None) == WHOLE
+
+
+def test_out_of_range_or_nonint_indices_fall_back_never_empty():
+    assert _scene_transcript_text(WHOLE, SEGS, [9, 10], None) == WHOLE          # all out of range
+    assert _scene_transcript_text(WHOLE, SEGS, ["1", "2"], None) == WHOLE       # 1-based strings, rejected
+    assert _scene_transcript_text(WHOLE, SEGS, [-1], None) == WHOLE            # negative
+    # partial validity: keep what's in range
+    assert _scene_transcript_text(WHOLE, SEGS, [3, 99], None) == "one occupant"