Deploy v1 — single-Docker FastAPI + Next.js + RAG + voice + faithfulness

backend/faithfulness.py

@@ -269,8 +269,13 @@ async def check_faithfulness(
         verdict.reasons.extend([f"unsupported_number: {n}" for n in bad_nums])
         verdict.unsupported_claims.extend(bad_nums)
 
-    # Gate 4 — LLM judge (only if previous gates passed — saves token cost on obvious failures)
-    if verdict.passed and run_llm_judge:
+    # Gate 4 — LLM judge (only if previous gates passed — saves token cost on
+    # obvious failures). Also SKIP when retrieval was strongly grounded: top
+    # chunk cosine > HIGH_CONFIDENCE_FLOOR means hallucination risk is low and
+    # the 1-2s Groq round-trip rarely adds value. Cuts ~60% of judge calls.
+    HIGH_CONFIDENCE_FLOOR = 0.50
+    top_score = max((c.score for c in chunks), default=0.0) if chunks else 0.0
+    if verdict.passed and run_llm_judge and top_score < HIGH_CONFIDENCE_FLOOR:
         ok4, unsupported = await _gate_llm_judge(reply, chunks)
         if not ok4:
             verdict.passed = False

rag/vectors/chroma.sqlite3

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5cc501a49d70adbf2c71cd93612b7a1f32b48fa6c7e33e57150769cd1fad9032
+oid sha256:562b5b6316b04ffaf466b02084daa347563ce1746f1f64e6644e738b0206e1d8
 size 87347200