Update utils.py

utils.py

@@ -61,6 +61,115 @@ def extract_first_json_obj(s: str) -> Optional[Dict[str, Any]]:
         return None
 
 
+# ----------------------------
+# Extra robustness: remove stray unquoted fields (e.g., `confidence: 0.75`)
+# that sometimes appear outside JSON strings due to generation glitches.
+# ----------------------------
+def _remove_unquoted_confidence_field(json_text: str) -> str:
+    """
+    Removes an unquoted trailing field like `confidence: 0.75` that appears
+    outside strings in otherwise-valid JSON output. This is a targeted fix
+    for common LLM glitches and intentionally conservative (only triggers
+    when we are NOT inside a quoted string).
+    """
+    out_chars: List[str] = []
+    i = 0
+    in_str = False
+    esc = False
+
+    def _pop_trailing_ws_and_optional_comma():
+        # remove trailing whitespace
+        while out_chars and out_chars[-1].isspace():
+            out_chars.pop()
+        # remove trailing comma (and whitespace before it)
+        if out_chars and out_chars[-1] == ",":
+            out_chars.pop()
+            while out_chars and out_chars[-1].isspace():
+                out_chars.pop()
+
+    while i < len(json_text):
+        ch = json_text[i]
+
+        if in_str:
+            out_chars.append(ch)
+            if esc:
+                esc = False
+            elif ch == "\\":  # escape
+                esc = True
+            elif ch == '"':
+                in_str = False
+            i += 1
+            continue
+
+        if ch == '"':
+            in_str = True
+            out_chars.append(ch)
+            i += 1
+            continue
+
+        # Detect an unquoted `confidence: <number>` outside strings.
+        # Only remove if followed by a number and then a delimiter (`,` or `}`).
+        if json_text.startswith("confidence", i):
+            j = i + len("confidence")
+            while j < len(json_text) and json_text[j].isspace():
+                j += 1
+            if j < len(json_text) and json_text[j] == ":":
+                j += 1
+                while j < len(json_text) and json_text[j].isspace():
+                    j += 1
+
+                # parse a simple number
+                if j < len(json_text) and json_text[j] in "+-":
+                    j += 1
+                has_digit = False
+                while j < len(json_text) and json_text[j].isdigit():
+                    has_digit = True
+                    j += 1
+                if j < len(json_text) and json_text[j] == ".":
+                    j += 1
+                    while j < len(json_text) and json_text[j].isdigit():
+                        has_digit = True
+                        j += 1
+
+                if has_digit:
+                    k = j
+                    while k < len(json_text) and json_text[k].isspace():
+                        k += 1
+                    if k < len(json_text) and json_text[k] in {",", "}"}:
+                        _pop_trailing_ws_and_optional_comma()
+                        i = k  # keep delimiter
+                        continue
+
+        out_chars.append(ch)
+        i += 1
+
+    return "".join(out_chars)
+
+
+def extract_json_obj_robust(s: str) -> Optional[Dict[str, Any]]:
+    """
+    Extract and parse the first JSON object from a model output string.
+
+    - Cuts at the first complete `{...}` (brace-balanced while respecting strings).
+    - Repairs a common glitch: an unquoted trailing `confidence: <num>`.
+    - Returns a dict if parsing succeeds, else None.
+    """
+    cut = stop_at_complete_json(s) or s
+    start = cut.find("{")
+    end = cut.rfind("}")
+    if start == -1 or end == -1 or end <= start:
+        return None
+
+    cand = cut[start : end + 1].strip()
+    cand = cand.replace("```json", "").replace("```", "").strip()
+    cand = _remove_unquoted_confidence_field(cand)
+
+    try:
+        return json.loads(cand)
+    except Exception:
+        return None
+
+
 # ----------------------------
 # Post-processing: remove template sentence
 # ----------------------------