fix(api): anchor JSON extraction on brackets in parse_json_markdown (#41959)

api/libs/json_in_md_parser.py

@@ -6,23 +6,35 @@ from core.llm_generator.output_parser.errors import OutputParserError
 def parse_json_markdown(json_string: str):
     # Get json from the backticks/braces
     json_string = json_string.strip()
-    starts = ["```json", "```", "``", "`", "{", "["]
-    ends = ["```", "``", "`", "}", "]"]
+    parsed: dict = {}
+
+    # Anchor on the JSON brackets themselves: from the first "{" or "[" to the
+    # last "}" or "]". This works whether or not the JSON is wrapped in code
+    # fences, and - unlike marker-priority search - is not confused by
+    # backticks inside JSON string values or in surrounding prose.
+    start_candidates = [i for i in (json_string.find("{"), json_string.find("[")) if i != -1]
+    if start_candidates:
+        start_index = min(start_candidates)
+        end_index = max(json_string.rfind("}"), json_string.rfind("]"))
+        if end_index != -1 and start_index < end_index:
+            end_index += 1
+            extracted_content = json_string[start_index:end_index].strip()
+            return json.loads(extracted_content)
+
+    # Fallback for fenced content without brackets (e.g. a fenced scalar).
+    starts = ["```json", "```", "``", "`"]
+    ends = ["```", "``", "`"]
     end_index = -1
     start_index = 0
-    parsed: dict = {}
     for s in starts:
         start_index = json_string.find(s)
         if start_index != -1:
-            if json_string[start_index] not in ("{", "["):
-                start_index += len(s)
+            start_index += len(s)
             break
     if start_index != -1:
         for e in ends:
             end_index = json_string.rfind(e, start_index)
             if end_index != -1:
-                if json_string[end_index] in ("}", "]"):
-                    end_index += 1
                 break
     if start_index != -1 and end_index != -1 and start_index < end_index:
         extracted_content = json_string[start_index:end_index].strip()

api/tests/unit_tests/libs/test_json_in_md_parser.py

@@ -107,3 +107,20 @@ def test_parse_and_check_json_markdown_handles_think_fenced_and_raw_variants():
     for src in cases:
         obj = parse_and_check_json_markdown(src, ["keywords", "category_id", "category_name"])
         assert obj == expected
+
+
+def test_parse_json_markdown_backtick_inside_string_value():
+    """Backticks inside JSON string values must not be mistaken for code fences."""
+    src = '{"code": "use `print` function", "n": 1}'
+    assert parse_json_markdown(src) == {"code": "use `print` function", "n": 1}
+
+
+def test_parse_json_markdown_backtick_in_surrounding_prose():
+    """Backticks in prose before the JSON must not break extraction."""
+    src = 'Here is `the` result: {"a": 1}'
+    assert parse_json_markdown(src) == {"a": 1}
+
+
+def test_parse_json_markdown_fenced_scalar_still_supported():
+    """Fenced content without brackets still parses via the fence fallback."""
+    assert parse_json_markdown('```json\n"hello"\n```') == "hello"