Sync from GitHub via hub-sync

llm.py

@@ -307,10 +307,31 @@ def _loads(s: str):
     return None
 
 
+def _scan_json_values(text: str) -> list:
+    """Walk the string and collect every top-level JSON value. Handles models
+    that emit several values with no array wrapper — e.g. MiniCPM-V on image
+    input returns `{...} {...} {...}` (space-separated objects, no brackets) —
+    and ignores junk between/around them (stray quotes, prose, trailing `"}`)."""
+    dec = json.JSONDecoder()
+    out, i, n = [], 0, len(text)
+    while i < n:
+        if text[i] in "{[":
+            try:
+                val, end = dec.raw_decode(text, i)
+                out.append(val)
+                i = end
+                continue
+            except ValueError:
+                pass
+        i += 1
+    return out
+
+
 def extract_json(text: str):
     """
-    Pull the first JSON object/array out of a model reply. Returns the parsed
-    object or None. Callers should handle None (skip card / use fallback grade).
+    Pull JSON out of a model reply. Returns the parsed object/array, a list when
+    the model emitted several values without an array wrapper, or None. Callers
+    must handle None (skip card / use fallback grade).
     """
     text = _strip_think(text.strip())
     # strip ```json fences if present
@@ -318,6 +339,19 @@ def extract_json(text: str):
     data = _loads(text)
     if data is not None:
         return data
+    # The whole text didn't parse — commonly because the model concatenated
+    # multiple JSON values (objects and/or arrays). Collect them all and flatten
+    # to a single list so callers expecting an array still work.
+    values = _scan_json_values(text)
+    if len(values) == 1:
+        return values[0]
+    if values:
+        flat: list = []
+        for v in values:
+            flat.extend(v) if isinstance(v, list) else flat.append(v)
+        return flat
+    # Last resort: a single object/array embedded in prose and/or over-escaped
+    # (\" / \n) — the plain scan above can't read that, but _loads can.
     match = re.search(r"(\[.*\]|\{.*\})", text, re.DOTALL)
     if match:
         return _loads(match.group(1))

test_extract_json.py

@@ -61,6 +61,28 @@ def test_overescaped_embedded_in_prose():
     print("ok  over-escaped object found inside prose")
 
 
+def test_concatenated_objects_no_array_wrapper():
+    # MiniCPM-V on image input returns several objects space-separated with no
+    # [ ] wrapper (+ a stray trailing quote) — verbatim shape from the OCR path.
+    raw = ('{"question":"Where do the light reactions occur?","answer":"Thylakoid '
+           'membranes.","topic":"Photosynthesis","difficulty":2} '
+           '{"question":"Where does the Calvin cycle take place?","answer":"The '
+           'stroma.","topic":"Photosynthesis","difficulty":1}"}')
+    data = llm.extract_json(raw)
+    assert isinstance(data, list) and len(data) == 2, data
+    assert data[0]["answer"] == "Thylakoid membranes."
+    assert data[1]["topic"] == "Photosynthesis"
+    print("ok  concatenated objects (no array wrapper) -> list")
+
+
+def test_concatenated_arrays_flattened():
+    raw = '[{"question":"A","answer":"a","topic":"T"}] [{"question":"B","answer":"b","topic":"T"}]'
+    data = llm.extract_json(raw)
+    assert isinstance(data, list) and len(data) == 2, data
+    assert [d["question"] for d in data] == ["A", "B"]
+    print("ok  concatenated arrays flattened -> single list")
+
+
 def test_valid_json_unaffected_regression():
     # Already-valid JSON (incl. a legitimately escaped quote inside a string)
     # parses on the first try and never hits the un-escape fallback.
@@ -76,5 +98,7 @@ if __name__ == "__main__":
     test_fully_escaped_with_newlines()
     test_overescaped_with_think_and_fence()
     test_overescaped_embedded_in_prose()
+    test_concatenated_objects_no_array_wrapper()
+    test_concatenated_arrays_flattened()
     test_valid_json_unaffected_regression()
     print("\nAll NAH-50 extract_json robustness tests passed.")