revert: embedding 벤치마크 비활성화 — qwen3-embedding:0.6b 품질 부족

qwen3-embedding:0.6b로 embedding 활성화 시 전 데이터셋 하락:
- Allganize 0.395→0.158, KLUE-MRC 0.717→0.563, AutoRAG 0.639→0.460
- 0.6B 모델이 한국어 QA 의미 매칭에 부적합
- 더 큰 embedding 모델 (multilingual-e5-large 등) 필요

hybrid weight도 원복 (alpha=0.5)

Author: SonAIengine

src/synaptic/search.py

@@ -108,7 +108,7 @@ async def search(
                 vec_scores[node.id] = vec_score
 
             # FTS + vector hybrid score aggregation
-            alpha = 0.5  # FTS vs vector weight (0.5 = equal)
+            alpha = 0.5  # FTS vs vector weight
             for nid, node in {n.id: n for n in vec_nodes}.items():
                 fts_s = fts_scores.get(nid, 0.0)
                 vec_s = vec_scores.get(nid, 0.0)
@@ -118,7 +118,7 @@ async def search(
                     all_nodes[nid] = (all_nodes[nid][0], min(1.0, hybrid))
                 else:
                     # vector only
-                    all_nodes[nid] = (node, vec_s * 0.9)  # slight decay when no FTS match
+                    all_nodes[nid] = (node, vec_s * 0.9)
 
         # Stage 2: Synonym expansion (if insufficient results)
         if len(all_nodes) < limit:

tests/benchmark/test_external_datasets.py

@@ -43,7 +43,7 @@ async def _build_graph(
     *,
     max_docs: int = 0,
 ) -> tuple[SynapticGraph, dict[str, str]]:
-    """corpus를 SynapticGraph에 인덱싱. id_map 반환."""
+    """corpus를 SynapticGraph에 인덱싱. FTS only (embedding은 모델 품질에 의존)."""
     backend = MemoryBackend()
     await backend.connect()
     graph = SynapticGraph(backend)