optimized prompt

app.py

@@ -653,18 +653,32 @@ def run_query(requirements: str):
     else:
         scored, _ = _score_items(filtered, keyword, debug=False)
         if scored:
-            scored.sort(key=lambda t: t[1], reverse=True)
-            filtered = [c for c, _, _ in scored]
+            # Blend LLM relevance with normalized prize to unify ranking
+            # prize_norm in [0,1] across the current candidate set
+            prizes = [c.prize for c, _, _ in scored]
+            pmin = min(prizes) if prizes else 0.0
+            pmax = max(prizes) if prizes else 0.0
+            denom = (pmax - pmin) if (pmax - pmin) > 0 else 1.0
+            def prize_norm(v: float) -> float:
+                return max(0.0, min(1.0, (v - pmin) / denom))
+
+            alpha = 0.75  # relevance weight; 25% weight to prize
+            ranked = []
+            for c, s, r in scored:
+                cs = alpha * float(s) + (1 - alpha) * prize_norm(c.prize)
+                ranked.append((c, cs, r))
+            ranked.sort(key=lambda t: t[1], reverse=True)
+            filtered = [c for c, _, _ in ranked]
     # If MCP returned items but filtering removed all (e.g., many items missing prize), show unfiltered
     if not filtered and items:
         filtered = items
         status = f"{status} (no matches; showing unfiltered)"
-    plan_text, _ = _generate_plan_fixed(scored, keyword, days_ahead, debug=False) if ok and scored else ("", "")
+    plan_text, _ = _generate_plan_fixed(ranked, keyword, days_ahead, debug=False) if ok and 'ranked' in locals() and ranked else ("", "")
 
     # Build a map from id to (score, reason) for star display
     id_to_score_reason: dict[str, tuple[float, str]] = {}
-    if ok:
-        for c, s, r in (scored if 'scored' in locals() else []):
+    if ok and 'ranked' in locals():
+        for c, s, r in ranked:
             id_to_score_reason[c.id] = (s, r)
 
     rows = []