From 95c95d95ab3afb4a134f4abeba08764078480eb5 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?LLM=E4=B8=9A=E5=8A=A1=E6=95=B0=E6=8D=AE=E4=B8=93=E5=AE=B6?=
 <266125204+alibababye@users.noreply.github.com>
Date: Tue, 23 Jun 2026 01:23:11 +0800
Subject: [PATCH] fix: parse embedded ranking JSON

Handle nested ranking JSON objects when an LLM wraps the response in explanatory prose. This keeps criterion scores from silently falling back to defaults.
---
 backend/app/services/ranking_service.py |  2 +-
 backend/tests/test_ranking_variance.py  | 31 ++++++++++++++++++++++++-
 2 files changed, 31 insertions(+), 2 deletions(-)

diff --git a/backend/app/services/ranking_service.py b/backend/app/services/ranking_service.py
index 6d030fc..d1891a7 100644
--- a/backend/app/services/ranking_service.py
+++ b/backend/app/services/ranking_service.py
@@ -289,7 +289,7 @@ def _parse_llm_response(
         except json.JSONDecodeError:
             # Try to find JSON object in text
             import re
-            match = re.search(r'\{[^{}]*\}', response_text, re.DOTALL)
+            match = re.search(r'\{[\s\S]*\}', response_text)
             if match:
                 try:
                     data = json.loads(match.group())
diff --git a/backend/tests/test_ranking_variance.py b/backend/tests/test_ranking_variance.py
index 0c426c0..405b2eb 100644
--- a/backend/tests/test_ranking_variance.py
+++ b/backend/tests/test_ranking_variance.py
@@ -10,7 +10,7 @@
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
 
 from app.models.idea import IdeaCandidate, DraftPlan, RiskItem, ExperimentSpec
-from app.services.ranking_service import RankingService
+from app.services.ranking_service import PAPER_TYPE_WEIGHTS, RankingService
 from datetime import datetime
 
 
@@ -93,6 +93,34 @@ def _make_candidate(idx: int, session_id: str = "test_session") -> IdeaCandidate
     )
 
 
+def test_parse_embedded_nested_json_response():
+    """LLM ranking JSON may be wrapped in explanatory prose."""
+    service = RankingService()
+    response = """
+    Here is the ranking result:
+    {
+      "novelty": {"score": 9.0, "rationale": "Fresh direction"},
+      "feasibility": {"score": 8.0, "rationale": "Buildable"},
+      "impact": {"score": 7.0, "rationale": "Useful"},
+      "clarity": {"score": 6.0, "rationale": "Clear enough"},
+      "risk": {"score": 5.0, "rationale": "Manageable"},
+      "alignment": {"score": 4.0, "rationale": "Partial fit"},
+      "referenceSupport": {"score": 3.0, "rationale": "Sparse refs"},
+      "experimentSpecificity": {"score": 2.0, "rationale": "Needs detail"},
+      "overallRationale": "Good candidate",
+      "confidence": 0.9
+    }
+    """
+
+    result = service._parse_llm_response("cand_embedded", response, PAPER_TYPE_WEIGHTS["default"])
+
+    assert result.criteria["novelty"].value == 9.0
+    assert result.criteria["experimentSpecificity"].value == 2.0
+    assert result.overallRationale == "Good candidate"
+    assert result.confidence == 0.9
+    print("PASS: test_parse_embedded_nested_json_response")
+
+
 def test_heuristic_scores_are_not_all_equal():
     """Scores must vary across 6 distinct candidates (heuristic fallback)."""
     service = RankingService()
@@ -197,6 +225,7 @@ def test_score_breakdown_structure():
 
 
 if __name__ == "__main__":
+    test_parse_embedded_nested_json_response()
     test_overall_score_uses_all_8_criteria()
     test_score_breakdown_structure()
     test_heuristic_scores_are_not_all_equal()