Relax verification for text documents - support flexible claim matching

- Verification now accepts: high similarity + neutral NLI, moderate similarity
+ entailment, or very high similarity alone
- Lowered default similarity threshold from 0.75 to 0.6 for paraphrased text
- Lowered firewall threshold from 0.8 to 0.6 for text document queries
- Fixes issue where valid PDF answers were marked as unsupported

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

config/settings.py

@@ -22,11 +22,11 @@ GROQ_API_KEY = os.getenv("GROQ_API_KEY", "")
 
 # Semantic similarity threshold (theta_sim)
 # Claims with similarity below this are considered unsupported
-SIMILARITY_THRESHOLD = 0.75
+SIMILARITY_THRESHOLD = 0.6
 
 # Firewall threshold (tau)
 # Responses with SupportRatio below this trigger regeneration
-FIREWALL_THRESHOLD = 0.8
+FIREWALL_THRESHOLD = 0.6
 
 # =============================================================================
 # DOCUMENT INGESTION PARAMETERS

core/verifier.py

@@ -282,10 +282,12 @@ class ClaimVerifier:
             hypothesis=claim.text
         )
 
-        # Apply verification rule
+        # Apply verification rule:
+        # Supported if EITHER high similarity OR entailment confirms it
         is_supported = (
-            similarity_score >= self.similarity_threshold and
-            entailment_label == 'ENTAILED'
+            (similarity_score >= self.similarity_threshold and entailment_label in ('ENTAILED', 'NEUTRAL')) or
+            (similarity_score >= 0.5 and entailment_label == 'ENTAILED') or
+            (similarity_score >= 0.85)
         )
 
         # Update claim object