fix threashold

api.py

@@ -223,6 +223,39 @@ def _normalize_image_path(raw_path: str) -> str:
 # ---------------------------------------------------------------------------
 
 FAISS_CONFIDENCE_THRESHOLD = float(os.getenv("FAISS_CONFIDENCE_THRESHOLD", "0.82"))
+FAISS_AMBIGUITY_MARGIN = float(os.getenv("FAISS_AMBIGUITY_MARGIN", "0.015"))
+RRF_AMBIGUITY_MARGIN = float(os.getenv("RRF_AMBIGUITY_MARGIN", "0.0015"))
+FORCE_OPENAI_FALLBACK = os.getenv("FORCE_OPENAI_FALLBACK", "0").strip().lower() in {
+    "1", "true", "yes", "on"
+}
+
+
+def _should_trigger_gpt_fallback(top_score: float, results: list[tuple[str, float, list]]) -> tuple[bool, str]:
+    """Decide whether GPT vision fallback should run.
+
+    Triggers on low FAISS confidence, explicit force flag, or very ambiguous top-vs-second gap.
+    """
+    if FORCE_OPENAI_FALLBACK:
+        return True, "forced_by_env"
+
+    if top_score < FAISS_CONFIDENCE_THRESHOLD:
+        return True, "low_top_score"
+
+    if len(results) < 2:
+        return False, "single_result"
+
+    top_result_score = float(results[0][1])
+    second_result_score = float(results[1][1])
+    gap = max(0.0, top_result_score - second_result_score)
+    rrf_like = top_result_score <= 0.1 and second_result_score <= 0.1
+
+    if rrf_like and gap < RRF_AMBIGUITY_MARGIN:
+        return True, "ambiguous_rrf_gap"
+
+    if (not rrf_like) and gap < FAISS_AMBIGUITY_MARGIN:
+        return True, "ambiguous_similarity_gap"
+
+    return False, "high_confidence"
 
 
 def _gpt_vision_identify_plant(
@@ -1291,11 +1324,13 @@ async def search_similar(
         gpt_job_status: dict[str, Any] | None = None
         gpt_fallback_attempted = False
         gpt_fallback_reason = "not_attempted"
-        if top_planclef_score < FAISS_CONFIDENCE_THRESHOLD and api_key:
+        should_trigger_gpt, gpt_trigger_basis = _should_trigger_gpt_fallback(top_planclef_score, results)
+        if should_trigger_gpt and api_key:
             gpt_fallback_attempted = True
             logger.info(
-                f"FAISS top score {top_planclef_score:.4f} < threshold {FAISS_CONFIDENCE_THRESHOLD} "
-                f"— activating GPT-4o vision fallback"
+                "Activating GPT-4o vision fallback: "
+                f"basis={gpt_trigger_basis}, top_planclef_score={top_planclef_score:.4f}, "
+                f"threshold={FAISS_CONFIDENCE_THRESHOLD}"
             )
             fallback_candidates = [species for species, _, _ in results[:12]]
             gpt_species, gpt_fallback_reason = _gpt_vision_identify_plant(
@@ -1314,7 +1349,7 @@ async def search_similar(
                 results = results[:k]
             else:
                 logger.info(f"GPT fallback attempted but no species accepted: {gpt_fallback_reason}")
-        elif top_planclef_score < FAISS_CONFIDENCE_THRESHOLD:
+        elif should_trigger_gpt:
             gpt_fallback_reason = "OPENAI_API_KEY missing"
 
     except RuntimeError as e:
@@ -1351,6 +1386,7 @@ async def search_similar(
             "gpt_fallback_attempted": gpt_fallback_attempted if 'gpt_fallback_attempted' in dir() else False,
             "gpt_fallback_used": gpt_species is not None if 'gpt_species' in dir() else False,
             "gpt_fallback_reason": gpt_fallback_reason if 'gpt_fallback_reason' in dir() else "not_attempted",
+            "gpt_trigger_basis": gpt_trigger_basis if 'gpt_trigger_basis' in dir() else "not_evaluated",
             "gpt_job_status": gpt_job_status if 'gpt_job_status' in dir() else None,
             "species_found": [species for species, _, _ in results],
             "scores": [float(score) for _, score, _ in results],

pwa-app/src/App.jsx

@@ -87,19 +87,10 @@ function normalizeSearchResults(rawResults) {
   const scores = sorted.map((item) => Number(item?.score ?? 0));
   const maxScore = Math.max(...scores);
 
-  // RRF scores are typically very small (around 0.01-0.03), so scale them for UI display.
-  const isRrfLike = maxScore <= 0.1;
-
   return sorted.map((item) => {
     const score = Number(item?.score ?? 0);
-    let displayScore = 0;
-
-    if (isRrfLike) {
-      // Keep relative proportions between RRF scores (no min-max stretching).
-      displayScore = maxScore > 0 ? score / maxScore : 0;
-    } else {
-      displayScore = Math.max(0, Math.min(1, score));
-    }
+    // Keep bars readable by always scaling against the best result in the current list.
+    const displayScore = maxScore > 0 ? Math.max(0, Math.min(1, score / maxScore)) : 0;
 
     return {
       ...item,