normalisasi rerank

app.py

@@ -59,7 +59,8 @@ def expand_query(query: str, num_variations: int = 3) -> str:
     try:
         # Panggil Gemini
         response = gemini_client.models.generate_content(
-            model="gemini-2.5-flash",
+            # model="gemini-2.5-flash",
+            model="gemini-robotics-er-1.5-preview",
             contents=prompt,
         )
 
@@ -190,22 +191,38 @@ def hybrid_search_no_gemini(query: str, match_count: int = 50):
     pairs = [(query, c["judul"] + " " + c["deskripsi"]) for c in candidates]
 
     try:
-        scores = reranker.predict(pairs)
-    except Exception as e:
-        print("Reranker error:", e)
-        return {"results": sorted(candidates, key=lambda x: x.get("similarity", 0), reverse=True)[:10]}
-
-    for c, s in zip(candidates, scores):
-        c["rerank_score"] = float(s)
-        rerank_prob = apply_sigmoid(float(s))
-        c["rerank_prob"] = rerank_prob
-
-        db_sim = c.get("similarity", 0.0)
-
-        c["final_score"] = (0.6 * rerank_prob) + (0.4 * db_sim)
-
-    candidates = sorted(candidates, key=lambda x: x["final_score"], reverse=True)
-    return {"results": candidates[:10]}
+            scores = reranker.predict(pairs)
+        except Exception as e:
+            print("Reranker error:", e)
+            return {"results": sorted(candidates, key=lambda x: x.get("similarity", 0), reverse=True)[:10]}
+    
+        # Ambil nilai min dan max dari skor RRF bawaan database
+        db_scores = [c.get("similarity", 0.0) for c in candidates]
+        db_min = min(db_scores) if db_scores else 0.0
+        db_max = max(db_scores) if db_scores else 1.0
+    
+        for c, s in zip(candidates, scores):
+            # 1. Probabilitas Reranker (0 sampai 1)
+            c["rerank_score"] = float(s)
+            rerank_prob = apply_sigmoid(float(s))
+            c["rerank_prob"] = rerank_prob
+            
+            # 2. Normalisasi Skor Database (RRF) ke 0 sampai 1
+            raw_db_sim = c.get("similarity", 0.0)
+            if db_max > db_min:
+                norm_db_sim = (raw_db_sim - db_min) / (db_max - db_min)
+            else:
+                norm_db_sim = 0.0
+                
+            c["norm_db_sim"] = norm_db_sim # Opsional: simpan untuk debugging
+    
+            # 3. Interpolasi yang SEIMBANG
+            c["final_score"] = (0.6 * rerank_prob) + (0.4 * norm_db_sim)
+    
+        # Urutkan berdasarkan final_score
+        candidates = sorted(candidates, key=lambda x: x["final_score"], reverse=True)
+        
+        return {"results": candidates[:10]}
 
 # ==========================================
 # CORE APPS