feat(api): use json_repair to fix invalid json while generating structured output (#18977)

When generating JSON schema using an LLM in the structured output feature,
models may occasionally return invalid JSON, which prevents clients from correctly 
parsing the response and can lead to UI breakage.

This commit addresses the issue by introducing `json_repair` to automatically 
fix invalid JSON strings returned by the LLM, ensuring smoother functionality 
and better client-side handling of structured outputs.


Co-authored-by: lizb <lizb@sugon.com>

api/core/llm_generator/llm_generator.py

@@ -3,6 +3,8 @@ import logging
 import re
 from typing import Optional, cast
 
+import json_repair
+
 from core.llm_generator.output_parser.rule_config_generator import RuleConfigGeneratorOutputParser
 from core.llm_generator.output_parser.suggested_questions_after_answer import SuggestedQuestionsAfterAnswerOutputParser
 from core.llm_generator.prompts import (
@@ -366,7 +368,20 @@ class LLMGenerator:
                 ),
             )
 
-            generated_json_schema = cast(str, response.message.content)
+            raw_content = response.message.content
+
+            if not isinstance(raw_content, str):
+                raise ValueError(f"LLM response content must be a string, got: {type(raw_content)}")
+
+            try:
+                parsed_content = json.loads(raw_content)
+            except json.JSONDecodeError:
+                parsed_content = json_repair.loads(raw_content)
+
+            if not isinstance(parsed_content, dict | list):
+                raise ValueError(f"Failed to parse structured output from llm: {raw_content}")
+
+            generated_json_schema = json.dumps(parsed_content, indent=2, ensure_ascii=False)
             return {"output": generated_json_schema, "error": ""}
 
         except InvokeError as e: