Add speaker labels

app/agents/claims.py

@@ -126,6 +126,40 @@ def run_claim_extractor(transcript: str) -> Dict[str, Any]:
     return {"claims": []}
 
 
+# =========================
+# Helper Functions
+# =========================
+
+def _find_speaker_for_claim(claim_text: str, segments: List[Dict]) -> str:
+    """
+    Find which speaker made a claim by matching claim text to transcript segments.
+    Uses fuzzy matching to handle slight variations in wording.
+    """
+    claim_words = set(claim_text.lower().split())
+    best_match_speaker = None
+    best_match_score = 0
+
+    for segment in segments:
+        segment_text = segment.get("text", "").lower()
+        segment_words = set(segment_text.split())
+
+        # Calculate overlap score (Jaccard similarity)
+        if segment_words:
+            intersection = claim_words.intersection(segment_words)
+            union = claim_words.union(segment_words)
+            score = len(intersection) / len(union) if union else 0
+
+            # Also check if claim is a substring (for exact matches)
+            if claim_text.lower() in segment_text or any(word in segment_text for word in claim_words if len(word) > 3):
+                score += 0.2  # Boost for substring matches
+
+            if score > best_match_score:
+                best_match_score = score
+                best_match_speaker = segment.get("speaker")
+
+    # Only return speaker if we have a reasonable confidence match
+    return best_match_speaker if best_match_score > 0.2 else None
+
 # =========================
 # Public: extract_claims (used by orchestrator)
 # =========================
@@ -146,10 +180,14 @@ def extract_claims(segments: List[Dict]) -> List[Claim]:
         text = (c.get("text") or "").strip()
         if not text:
             continue
+
+        # Map claim to speaker by finding which segment contains this text
+        speaker = _find_speaker_for_claim(text, segments)
+
         out.append(Claim(
             id=f"c{i}",
             text=text,
-            speaker=c.get("speaker"),
+            speaker=speaker or c.get("speaker"),
             segment_idx=0,  # TODO: map to true segment via start/end if available
             confidence=float(c.get("confidence", 0.6)),
         ))