Tighten MiniCPM5 JSON constraint parsing

app.py

@@ -1,6 +1,7 @@
 import json
 import math
 import os
+import ast
 import re
 from dataclasses import dataclass
 from functools import lru_cache
@@ -183,13 +184,29 @@ def normalize_constraints(raw: Dict[str, object]) -> Dict[str, object]:
 
 
 def extract_json_object(text: str) -> Optional[Dict[str, object]]:
-    match = re.search(r"\{.*\}", text or "", re.DOTALL)
+    cleaned = (text or "").strip()
+    fenced = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", cleaned, re.DOTALL | re.IGNORECASE)
+    if fenced:
+        cleaned = fenced.group(1)
+    match = re.search(r"\{.*\}", cleaned, re.DOTALL)
     if not match:
         return None
+    raw = match.group(0)
     try:
-        parsed = json.loads(match.group(0))
+        parsed = json.loads(raw)
     except json.JSONDecodeError:
-        return None
+        try:
+            parsed = ast.literal_eval(raw)
+        except (SyntaxError, ValueError):
+            normalized = (
+                raw.replace("True", "true")
+                .replace("False", "false")
+                .replace("None", "null")
+            )
+            try:
+                parsed = json.loads(normalized)
+            except json.JSONDecodeError:
+                return None
     return parsed if isinstance(parsed, dict) else None
 
 
@@ -223,25 +240,32 @@ def minicpm_parse_dispatch_notes(notes: str) -> Tuple[Dict[str, object], str]:
         fallback["source"] = "rule-fallback"
         return fallback, "MiniCPM5 local runtime is unavailable, so the deterministic parser handled the notes."
 
-    prompt = f"""You are a dispatch constraint parser for a small delivery route planner.
-Return only valid JSON with these keys:
-prefer_early_priority: boolean
-avoid_late_penalty: number between 0.5 and 10
-max_route_load: integer
-depot_start: minutes after midnight as integer
-soft_due_before: optional minutes after midnight
-boost_terms: short lowercase strings
-
-Dispatcher notes:
-{notes}
+    prompt = f"""<|im_start|>system
+You convert dispatcher notes into one compact JSON object.
+Return JSON only. No markdown. No explanation.
+Schema:
+{{
+  "prefer_early_priority": true,
+  "avoid_late_penalty": 2.0,
+  "max_route_load": 18,
+  "depot_start": 480,
+  "soft_due_before": 720,
+  "boost_terms": ["school", "medical", "fresh"]
+}}
+Use minutes after midnight for times. Omit soft_due_before if no lunch/noon constraint is present.
+<|im_end|>
+<|im_start|>user
+Dispatcher notes: {notes}
+<|im_end|>
+<|im_start|>assistant
 """
     try:
         result = llm(
             prompt,
             max_tokens=180,
-            temperature=0.1,
-            top_p=0.9,
-            stop=["\n\n"],
+            temperature=0.0,
+            top_p=1.0,
+            stop=["<|im_end|>", "\n\n\n"],
         )
         text = result["choices"][0]["text"]
         parsed = extract_json_object(text)