cc.ja.300.binのパスの参照先をfiles.jsonに一本化

api/search/engine.py

@@ -96,9 +96,11 @@ class SearchEngine:
             word_sim_enable=bool(search("word_sim.enable")),
             word_sim_alpha=float(search("word_sim.alpha")),
             word_sim_topk_k=int(search("word_sim.topk_k", 3)),
-            word_sim_rerank=(search("word_sim.rerank", "pair_avg") or "pair_avg").lower(),
+            word_sim_rerank=(
+                search("word_sim.rerank", "pair_avg") or "pair_avg"
+            ).lower(),
             query_subword_enable=bool(search("query_subword.enable")),
-            query_subword_path=search("query_subword.path"),
+            query_subword_path=files("embeddings.fasttext_bin"),
             query_subword_oov_weight=float(search("query_subword.oov_weight")),
             org_boost_exact=float(search("organization.boost.exact", 1.0)),
             org_boost_prefix=float(search("organization.boost.prefix", 0.7)),
@@ -132,7 +134,9 @@ class SearchEngine:
             log.warning(f"failed to load synonyms_cache: {e}")
 
         try:
-            custom_path = field_getter("config/search_model.json")("synonyms.sources.custom_json")
+            custom_path = field_getter("config/search_model.json")(
+                "synonyms.sources.custom_json"
+            )
             if custom_path and os.path.exists(custom_path):
                 with open(custom_path, encoding="utf-8") as f:
                     self.custom_synonyms = json.load(f)
@@ -143,8 +147,14 @@ class SearchEngine:
         with open(files("projects.projects_json"), encoding="utf-8") as f:
             self.projects = json.load(f)
         self.project_map = {p["projectId"]: p for p in self.projects}
-        self.org_norms = {p["projectId"]: normalize_text_for_org(p.get("organization") or "") for p in self.projects}
-        self.reading_norms = {p["projectId"]: normalize_text_for_org(p.get("reading") or "") for p in self.projects}
+        self.org_norms = {
+            p["projectId"]: normalize_text_for_org(p.get("organization") or "")
+            for p in self.projects
+        }
+        self.reading_norms = {
+            p["projectId"]: normalize_text_for_org(p.get("reading") or "")
+            for p in self.projects
+        }
 
         # BM25F assets
         with open(files("bm25.bm25_meta"), encoding="utf-8") as f:
@@ -169,7 +179,11 @@ class SearchEngine:
         # doc_vectors.npy は topk 方式では不要
 
         # fastText OOV
-        if self.cfg.query_subword_enable and self.cfg.query_subword_path and os.path.exists(self.cfg.query_subword_path):
+        if (
+            self.cfg.query_subword_enable
+            and self.cfg.query_subword_path
+            and os.path.exists(self.cfg.query_subword_path)
+        ):
             try:
                 import fasttext
 
@@ -246,7 +260,7 @@ class SearchEngine:
                 denom_sum = 0.0
                 num_sum = 0.0
                 for field, weight in fw.items():
-                    fobj = (fields.get(field) or {})
+                    fobj = fields.get(field) or {}
                     tf = float((fobj.get("tf") or {}).get(t, 0))
                     if tf <= 0.0:
                         continue
@@ -314,8 +328,8 @@ class SearchEngine:
             fields = d.get("fields") or {}
             doc_terms = set()
             for fname in self.cfg.target_fields:
-                fobj = (fields.get(fname) or {})
-                tf = (fobj.get("tf") or {})
+                fobj = fields.get(fname) or {}
+                tf = fobj.get("tf") or {}
                 doc_terms.update(tf.keys())
             if not doc_terms:
                 sims_all[i] = 0.0
@@ -371,8 +385,8 @@ class SearchEngine:
             fields = d.get("fields") or {}
             doc_terms = set()
             for fname in self.cfg.target_fields:
-                fobj = (fields.get(fname) or {})
-                tf = (fobj.get("tf") or {})
+                fobj = fields.get(fname) or {}
+                tf = fobj.get("tf") or {}
                 doc_terms.update(tf.keys())
             if not doc_terms:
                 sims_all[i] = 0.0
@@ -442,9 +456,15 @@ class SearchEngine:
         # Relative threshold anchored to the top fused score
         score_with_boost = fused_filter + boost
         top = float(np.max(score_with_boost)) if score_with_boost.size > 0 else 0.0
-        rel_cut = top * float(self.cfg.fused_rel_top_ratio) if top > 0 else self.cfg.fused_min
+        rel_cut = (
+            top * float(self.cfg.fused_rel_top_ratio) if top > 0 else self.cfg.fused_min
+        )
         fused_cut = max(float(self.cfg.fused_min), rel_cut)
-        keep = ((bm25 >= self.cfg.bm25_min) | (ws_filter >= self.cfg.word_sim_min) | (score_with_boost >= self.cfg.fused_min)) & (score_with_boost >= fused_cut)
+        keep = (
+            (bm25 >= self.cfg.bm25_min)
+            | (ws_filter >= self.cfg.word_sim_min)
+            | (score_with_boost >= self.cfg.fused_min)
+        ) & (score_with_boost >= fused_cut)
         order = np.argsort(-score_with_boost)  # descending by fused
         selected_idx: List[int] = []
         for i in order:
@@ -454,7 +474,11 @@ class SearchEngine:
                 break
         # Single-step fallback: if zero, relax the relative cut and use absolute thresholds only
         if len(selected_idx) == 0:
-            keep2 = (bm25 >= self.cfg.bm25_min) | (ws_filter >= self.cfg.word_sim_min) | (score_with_boost >= self.cfg.fused_min)
+            keep2 = (
+                (bm25 >= self.cfg.bm25_min)
+                | (ws_filter >= self.cfg.word_sim_min)
+                | (score_with_boost >= self.cfg.fused_min)
+            )
             for i in order:
                 if keep2[i]:
                     selected_idx.append(int(i))
@@ -477,15 +501,19 @@ class SearchEngine:
         # build debug details for selected docs
         details = []
         for i in selected_idx:
-            details.append({
-                "projectId": ids[i],
-                "bm25": float(bm25[i]),
-                "ws_filter_topk": float(ws_filter[i]),
-                "ws_rerank_pairavg": float(ws_rerank[i]) if ws_rerank is not None else None,
-                "org_boost": float(boost[i]),
-                "fused_filter": float(fused_filter[i]),
-                "fused_final": float(final_scores[i]),
-            })
+            details.append(
+                {
+                    "projectId": ids[i],
+                    "bm25": float(bm25[i]),
+                    "ws_filter_topk": float(ws_filter[i]),
+                    "ws_rerank_pairavg": float(ws_rerank[i])
+                    if ws_rerank is not None
+                    else None,
+                    "org_boost": float(boost[i]),
+                    "fused_filter": float(fused_filter[i]),
+                    "fused_final": float(final_scores[i]),
+                }
+            )
         return pairs, {"details": details}
 
     def get_projects(self) -> List[Dict[str, Any]]:

config/search_model.json

@@ -51,7 +51,6 @@
   },
   "query_subword": {
     "enable": true,
-    "path": "resources/embeddings/cc.ja.300.bin",
     "oov_weight": 0.8,
     "cache_size": 50000
   },