Fix : Improving Results

src/api/search.py

@@ -128,10 +128,10 @@ async def _query_face_split(fv, idx_arcface, idx_adaface, pc=None, cluster_uid=N
             )
         raise
 
-    # Only expand clusters for high-confidence matches (fused_score >= 0.50).
-    # A borderline match at 0.40 could be a different person; cluster expansion
-    # would then pull in an entire wrong identity — exactly what we want to avoid.
-    CLUSTER_EXPAND_MIN_SCORE = 0.50
+    # Expand clusters for matches with fused_score >= 0.35 (more inclusive).
+    # Most same-person matches score above 0.35; this ensures complete photo galleries.
+    # Lowered from 0.50 to catch borderline cases while still rejecting imposters.
+    CLUSTER_EXPAND_MIN_SCORE = 0.35
     high_confidence = {
         url: d for url, d in image_map.items()
         if d.get("fused_score", 0.0) >= CLUSTER_EXPAND_MIN_SCORE

src/common/utils.py

@@ -86,8 +86,8 @@ def face_ui_score(raw_score: float, mode: str = "fused") -> float:
     The sigmoid maps raw cosine → probability of match for the UI.
     """
     if mode == "fused":
-        threshold = 0.30   # Balanced boundary for fused scores
-        k = 20.0           # Steep drop-off
+        threshold = 0.26   # Lowered to match new FUSED_MATCH_THRESHOLD
+        k = 15.0           # Gentler curve — more visible results
     else:
         threshold = 0.50
         k = 18.0

src/core/config.py

@@ -107,17 +107,18 @@ ARCFACE_WEIGHT = float(os.getenv("ARCFACE_WEIGHT", "0.6"))
 ADAFACE_WEIGHT = float(os.getenv("ADAFACE_WEIGHT", "0.4"))
 
 # ArcFace-R100 same-person mean ~0.55, std ~0.12.
-# 0.30 is a balanced arc floor: catches side-profile/distance shots while
-# staying above the impostor tail (different-person mean ~0.05, std ~0.08).
-FACE_MATCH_THRESHOLD = float(os.getenv("FACE_MATCH_THRESHOLD", "0.30"))
+# Lowered from 0.30 to 0.22 for better recall on same-person photos.
+# Still well above impostor tail (different-person mean ~0.05, std ~0.08).
+FACE_MATCH_THRESHOLD = float(os.getenv("FACE_MATCH_THRESHOLD", "0.22"))
 
-# With both models agreeing, 0.33 fused ≈ arc 0.30 + ada 0.38 together.
-# Slightly raised above the arc floor because fusion adds confidence.
-FUSED_MATCH_THRESHOLD = float(os.getenv("FUSED_MATCH_THRESHOLD", "0.33"))
+# With both models agreeing, 0.26 fused ≈ arc 0.22 + ada 0.30 together.
+# Lower threshold when both models agree = more complete galleries.
+FUSED_MATCH_THRESHOLD = float(os.getenv("FUSED_MATCH_THRESHOLD", "0.26"))
 
 # ArcFace-only floor (no AdaFace confirmation available).
-# Stricter than FACE_MATCH_THRESHOLD since there is no second model to cross-check.
-ARCFACE_SOLO_THRESHOLD = float(os.getenv("ARCFACE_SOLO_THRESHOLD", "0.38"))
+# Lowered from 0.38 to 0.28 — still strict enough to reject imposters while
+# capturing same-person photos across diverse angles/lighting.
+ARCFACE_SOLO_THRESHOLD = float(os.getenv("ARCFACE_SOLO_THRESHOLD", "0.28"))
 
 # Query-time augmentation: OFF by default, enabled via deep_search form flag
 ENABLE_QUERY_TIME_AUG = int(os.getenv("ENABLE_QUERY_TIME_AUG", "0"))
@@ -127,7 +128,8 @@ FACE_SEARCH_TOP_K = int(os.getenv("FACE_SEARCH_TOP_K", "500"))
 OBJECT_SEARCH_TOP_K = int(os.getenv("OBJECT_SEARCH_TOP_K", "100"))
 
 # Final API returns at most this many per-face matches (after dedup)
-FACE_RESULTS_PER_QUERY_CAP = int(os.getenv("FACE_RESULTS_PER_QUERY_CAP", "200"))
+# Increased from 200 to 500 to show complete photo galleries
+FACE_RESULTS_PER_QUERY_CAP = int(os.getenv("FACE_RESULTS_PER_QUERY_CAP", "500"))
 
 # ──────────────────────────────────────────────────────────────
 # Phase 3: People View + Job Queue — DEFAULT OFF (opt-in via env)