fix: harden ontology JSON generation

backend/app/api/graph.py

@@ -4,6 +4,7 @@
 """
 
 import os
+import re
 import traceback
 import threading
 from contextlib import ExitStack, nullcontext
@@ -24,6 +25,7 @@ from ..models.project import ProjectManager, ProjectStatus
 from ..services.simulation_manager import SimulationManager
 from ..services.simulation_runner import SimulationRunner, RunnerStatus
 from ..services.zep_graph_memory_updater import ZepGraphMemoryManager
+from ..utils.llm_client import LLMResponseError
 
 # 获取日志器
 logger = get_logger('mirofish.api')
@@ -288,6 +290,7 @@ def generate_ontology():
             }
         }
     """
+    project = None
     try:
         logger.info("=== 开始生成本体定义 ===")
         
@@ -388,12 +391,56 @@ def generate_ontology():
             }
         })
         
-    except Exception as e:
+    except Exception as error:
-        return jsonify({
+        provider_status = getattr(error, "status_code", None)
+        request_id = getattr(error, "request_id", None)
+
+        if isinstance(error, LLMResponseError):
+            public_error = str(error)
+            response_status = 502
+            logger.exception("LLM returned an unusable ontology response")
+        elif isinstance(provider_status, int):
+            public_error = f"LLM provider request failed (HTTP {provider_status})"
+            if request_id:
+                safe_request_id = re.sub(
+                    r"[^a-zA-Z0-9._:-]", "", str(request_id)
+                )[:128]
+                if safe_request_id:
+                    public_error += f" (request_id: {safe_request_id})"
+            response_status = 502
+            # Provider exception bodies may echo request content. Keep the
+            # server log useful without serializing the exception body.
+            logger.error(
+                "Ontology provider request failed: type=%s status=%s request_id=%s",
+                type(error).__name__,
+                provider_status,
+                request_id or "unknown",
+            )
+        else:
+            public_error = "Ontology generation failed; check the server logs"
+            response_status = 500
+            logger.exception("Unexpected ontology generation failure")
+
+        response_data = None
+        if project is not None:
+            project.status = ProjectStatus.FAILED
+            project.error = public_error
+            try:
+                ProjectManager.save_project(project)
+            except Exception:
+                logger.exception(
+                    "Failed to persist ontology failure for project %s",
+                    project.project_id,
+                )
+            response_data = {"project_id": project.project_id}
+
+        payload = {
             "success": False,
-            "error": str(e),
+            "error": public_error,
-            "traceback": traceback.format_exc()
+        }
-        }), 500
+        if response_data is not None:
+            payload["data"] = response_data
+        return jsonify(payload), response_status
 
 
 # ============== 接口2：构建图谱 ==============

backend/app/services/ontology_generator.py

@@ -235,7 +235,11 @@ class OntologyGenerator:
         result = self.llm_client.chat_json(
             messages=messages,
             temperature=0.3,
-            max_tokens=4096
+            # Structured ontology responses can exceed 4096 completion tokens,
+            # especially when a compatible provider counts hidden reasoning in
+            # the same budget. Let the provider use its model-specific limit.
+            max_tokens=None,
+            max_attempts=2,
         )
         
         # 验证和后处理

backend/app/utils/llm_client.py

@@ -4,6 +4,7 @@ LLM客户端封装
 """
 
 import json
+import logging
 import re
 from typing import Optional, Dict, Any, List
 from openai import OpenAI
@@ -12,6 +13,81 @@ from ..config import Config
 from .openai_chat_compat import create_chat_completion, extract_chat_completion_text
 
 
+logger = logging.getLogger(__name__)
+
+
+class LLMResponseError(ValueError):
+    """A safe, structured error for unusable model responses."""
+
+    def __init__(self, message: str, *, finish_reason: Optional[str] = None):
+        super().__init__(message)
+        self.finish_reason = finish_reason
+
+
+def _is_response_format_unsupported(error: Exception) -> bool:
+    """Detect an explicit provider rejection of JSON response_format."""
+
+    if getattr(error, "status_code", None) not in {400, 422}:
+        return False
+
+    body = getattr(error, "body", None)
+    if not isinstance(body, dict):
+        return False
+
+    details = body.get("error", body)
+    if not isinstance(details, dict):
+        return False
+
+    param = str(details.get("param") or "").strip().lower()
+    if param == "response_format" or param.startswith("response_format."):
+        return True
+
+    message = str(details.get("message") or "").lower()
+    if "response_format" not in message:
+        return False
+
+    code = str(details.get("code") or "").lower()
+    unsupported_codes = {
+        "unsupported_parameter",
+        "unsupported_value",
+        "unknown_parameter",
+        "invalid_parameter",
+    }
+    unsupported_phrases = (
+        "not support",
+        "unsupported",
+        "unknown parameter",
+        "unrecognized parameter",
+    )
+    return code in unsupported_codes or any(
+        phrase in message for phrase in unsupported_phrases
+    )
+
+
+def _clean_chat_text(content: str) -> str:
+    """Remove common reasoning wrappers and an outer Markdown JSON fence."""
+
+    cleaned = re.sub(r'<think>[\s\S]*?</think>', '', content).strip()
+    cleaned = cleaned.lstrip("\ufeff")
+    cleaned = re.sub(r'^```(?:json)?\s*\n?', '', cleaned, flags=re.IGNORECASE)
+    cleaned = re.sub(r'\n?```\s*$', '', cleaned)
+    return cleaned.strip()
+
+
+def _contains_additional_json_container(content: str) -> bool:
+    """Return True when trailing text embeds another JSON object or array."""
+
+    decoder = json.JSONDecoder()
+    for match in re.finditer(r"[\[{]", content):
+        try:
+            value, _ = decoder.raw_decode(content[match.start():])
+        except json.JSONDecodeError:
+            continue
+        if isinstance(value, (dict, list)):
+            return True
+    return False
+
+
 class LLMClient:
     """LLM客户端"""
     
@@ -33,11 +109,30 @@ class LLMClient:
             base_url=self.base_url
         )
 
+    def _create_completion(
+        self,
+        *,
+        messages: List[Dict[str, str]],
+        temperature: Optional[float],
+        max_tokens: Optional[int],
+        response_format: Optional[Dict[str, Any]],
+    ) -> Any:
+        """Send one raw Chat Completions request through the compatibility layer."""
+
+        return create_chat_completion(
+            self.client,
+            model=self.model,
+            messages=messages,
+            temperature=temperature,
+            max_tokens=max_tokens,
+            response_format=response_format,
+        )
+    
     def chat(
         self,
         messages: List[Dict[str, str]],
         temperature: float = 0.7,
-        max_tokens: int = 4096,
+        max_tokens: Optional[int] = 4096,
         response_format: Optional[Dict] = None
     ) -> str:
         """
@@ -52,24 +147,21 @@ class LLMClient:
         Returns:
             模型响应文本
         """
-        response = create_chat_completion(
+        response = self._create_completion(
-            self.client,
-            model=self.model,
             messages=messages,
             temperature=temperature,
             max_tokens=max_tokens,
             response_format=response_format,
         )
         content = extract_chat_completion_text(response)
-        # 部分模型（如MiniMax M2.5）会在content中包含<think>思考内容，需要移除
+        return _clean_chat_text(content)
-        content = re.sub(r'<think>[\s\S]*?</think>', '', content).strip()
-        return content
     
     def chat_json(
         self,
         messages: List[Dict[str, str]],
         temperature: float = 0.3,
-        max_tokens: int = 4096
+        max_tokens: Optional[int] = 4096,
+        max_attempts: int = 1,
     ) -> Dict[str, Any]:
         """
         发送聊天请求并返回JSON
@@ -78,24 +170,121 @@ class LLMClient:
             messages: 消息列表
             temperature: 温度参数
             max_tokens: 最大token数
+            max_attempts: 内容生成尝试次数（不含一次明确的JSON模式能力降级）
             
         Returns:
             解析后的JSON对象
         """
-        response = self.chat(
+        if max_attempts < 1:
-            messages=messages,
+            raise ValueError("max_attempts must be at least 1")
-            temperature=temperature,
+
-            max_tokens=max_tokens,
+        response_format: Optional[Dict[str, str]] = {"type": "json_object"}
-            response_format={"type": "json_object"}
+        request_max_tokens = max_tokens
-        )
+        last_error: Optional[LLMResponseError] = None
-        # 清理markdown代码块标记
+
-        cleaned_response = response.strip()
+        for attempt in range(1, max_attempts + 1):
-        cleaned_response = re.sub(r'^```(?:json)?\s*\n?', '', cleaned_response, flags=re.IGNORECASE)
+            # JSON-mode capability negotiation is separate from content
-        cleaned_response = re.sub(r'\n?```\s*$', '', cleaned_response)
+            # regeneration. An explicit response_format rejection may add one
-        cleaned_response = cleaned_response.strip()
+            # request, but it must not consume a content attempt.
+            while True:
+                try:
+                    response = self._create_completion(
+                        messages=messages,
+                        temperature=temperature,
+                        max_tokens=request_max_tokens,
+                        response_format=response_format,
+                    )
+                except Exception as error:
+                    if (
+                        response_format is not None
+                        and _is_response_format_unsupported(error)
+                    ):
+                        logger.warning(
+                            "LLM provider explicitly rejected response_format; "
+                            "retrying once with prompt-only JSON guidance"
+                        )
+                        response_format = None
+                        continue
+                    raise
+                break
+
+            try:
+                return self._parse_json_response(response)
+            except LLMResponseError as error:
+                last_error = error
+                if attempt >= max_attempts:
+                    raise
+
+                # A caller-supplied cap is the common cause of a partial JSON
+                # object. Omit it for the one bounded retry so the provider can
+                # use its model-specific output limit.
+                had_token_cap = request_max_tokens is not None
+                request_max_tokens = None
+                logger.warning(
+                    "LLM returned unusable JSON (finish_reason=%s); "
+                    "retrying content generation%s",
+                    error.finish_reason or "unknown",
+                    " without an output token cap" if had_token_cap else "",
+                )
+
+        if last_error is not None:  # pragma: no cover - defensive loop guard
+            raise last_error
+        raise LLMResponseError("LLM did not produce a JSON response")
+
+    @staticmethod
+    def _parse_json_response(response: Any) -> Dict[str, Any]:
+        choices = getattr(response, "choices", None) or []
+        if not choices:
+            raise LLMResponseError("LLM returned no choices")
+
+        choice = choices[0]
+        finish_reason = getattr(choice, "finish_reason", None)
+        if finish_reason == "length":
+            raise LLMResponseError(
+                "LLM JSON output was truncated at the token limit",
+                finish_reason=finish_reason,
+            )
+        if finish_reason not in {None, "stop"}:
+            raise LLMResponseError(
+                f"LLM JSON generation stopped unexpectedly ({finish_reason})",
+                finish_reason=finish_reason,
+            )
+
+        content = _clean_chat_text(extract_chat_completion_text(response))
+        if not content:
+            raise LLMResponseError(
+                "LLM returned empty JSON content",
+                finish_reason=finish_reason,
+            )
 
         try:
-            return json.loads(cleaned_response)
+            value = json.loads(content)
-        except json.JSONDecodeError:
+        except json.JSONDecodeError as strict_error:
-            raise ValueError(f"LLM返回的JSON格式无效: {cleaned_response}")
+            # Some compatible providers append a short explanation after an
+            # otherwise complete JSON object. Accept only an object decoded
+            # from the beginning; never repair or invent truncated JSON.
+            try:
+                value, end = json.JSONDecoder().raw_decode(content)
+            except json.JSONDecodeError:
+                raise LLMResponseError(
+                    "LLM returned invalid JSON "
+                    f"(line {strict_error.lineno}, column {strict_error.colno})",
+                    finish_reason=finish_reason,
+                ) from strict_error
 
+            trailing = content[end:].strip()
+            if trailing:
+                if _contains_additional_json_container(trailing):
+                    raise LLMResponseError(
+                        "LLM returned multiple JSON values",
+                        finish_reason=finish_reason,
+                    )
+                logger.warning("Ignoring text after a complete LLM JSON object")
+
+        if not isinstance(value, dict):
+            raise LLMResponseError(
+                "LLM JSON response must be a top-level JSON object",
+                finish_reason=finish_reason,
+            )
+
+        return value

backend/tests/test_llm_json_responses.py

@@ -0,0 +1,259 @@
+from types import SimpleNamespace
+
+import pytest
+
+from app.utils.llm_client import LLMClient, LLMResponseError
+
+
+class CompletionSequence:
+    def __init__(self, *results):
+        self.results = list(results)
+        self.calls = []
+
+    def create(self, **kwargs):
+        self.calls.append(kwargs)
+        result = self.results.pop(0)
+        if isinstance(result, Exception):
+            raise result
+        return result
+
+
+class ProviderError(RuntimeError):
+    def __init__(self, *, status_code, body):
+        super().__init__(body.get("error", {}).get("message", "provider error"))
+        self.status_code = status_code
+        self.body = body
+
+
+def _response(content, *, finish_reason="stop", include_choice=True):
+    choices = []
+    if include_choice:
+        choices.append(
+            SimpleNamespace(
+                finish_reason=finish_reason,
+                message=SimpleNamespace(content=content),
+            )
+        )
+    return SimpleNamespace(choices=choices)
+
+
+def _client_for(sequence):
+    client = object.__new__(LLMClient)
+    client.model = "compatible-model"
+    client.client = SimpleNamespace(
+        chat=SimpleNamespace(completions=sequence)
+    )
+    return client
+
+
+def test_chat_json_retries_truncated_completion_without_token_cap():
+    sequence = CompletionSequence(
+        _response('{"items": [', finish_reason="length"),
+        _response('{"items": [1, 2]}'),
+    )
+    client = _client_for(sequence)
+
+    result = client.chat_json(
+        messages=[{"role": "user", "content": "Return JSON"}],
+        max_tokens=4096,
+        max_attempts=2,
+    )
+
+    assert result == {"items": [1, 2]}
+    assert sequence.calls[0]["max_tokens"] == 4096
+    assert "max_tokens" not in sequence.calls[1]
+
+
+def test_chat_json_omits_token_cap_when_requested():
+    sequence = CompletionSequence(_response('{"ok": true}'))
+    client = _client_for(sequence)
+
+    assert client.chat_json(
+        messages=[{"role": "user", "content": "Return JSON"}],
+        max_tokens=None,
+    ) == {"ok": True}
+    assert "max_tokens" not in sequence.calls[0]
+
+
+def test_chat_json_retries_empty_content_once():
+    sequence = CompletionSequence(
+        _response(None),
+        _response('{"ok": true}'),
+    )
+    client = _client_for(sequence)
+
+    result = client.chat_json(
+        messages=[{"role": "user", "content": "Return JSON"}],
+        max_attempts=2,
+    )
+
+    assert result == {"ok": True}
+    assert len(sequence.calls) == 2
+
+
+def test_chat_json_accepts_complete_object_before_trailing_text():
+    sequence = CompletionSequence(
+        _response('{"ok": true}\nThis object is ready to use.')
+    )
+    client = _client_for(sequence)
+
+    assert client.chat_json(
+        messages=[{"role": "user", "content": "Return JSON"}],
+    ) == {"ok": True}
+
+
+@pytest.mark.parametrize(
+    "content",
+    [
+        '{"status": "draft"}\n{"status": "complete"}',
+        '{"status": "draft"}\n```json\n{"status": "complete"}\n```',
+        '{"status": "draft"}\nExplanation first.\n{"status": "complete"}',
+    ],
+)
+def test_chat_json_rejects_a_second_json_document(content):
+    sequence = CompletionSequence(_response(content))
+    client = _client_for(sequence)
+
+    with pytest.raises(LLMResponseError, match="multiple JSON"):
+        client.chat_json(
+            messages=[{"role": "user", "content": "Return JSON"}],
+        )
+
+
+def test_chat_json_rejects_top_level_array():
+    sequence = CompletionSequence(_response('[{"ok": true}]'))
+    client = _client_for(sequence)
+
+    with pytest.raises(LLMResponseError, match="JSON object"):
+        client.chat_json(
+            messages=[{"role": "user", "content": "Return JSON"}],
+            max_attempts=1,
+        )
+
+
+def test_chat_json_error_does_not_echo_partial_model_output():
+    partial = '{"private_source_text": "SENTINEL-SHOULD-NOT-LEAK"'
+    sequence = CompletionSequence(
+        _response(partial, finish_reason="length"),
+        _response(partial, finish_reason="length"),
+    )
+    client = _client_for(sequence)
+
+    with pytest.raises(LLMResponseError) as captured:
+        client.chat_json(
+            messages=[{"role": "user", "content": "Return JSON"}],
+            max_attempts=2,
+        )
+
+    assert "SENTINEL-SHOULD-NOT-LEAK" not in str(captured.value)
+    assert captured.value.finish_reason == "length"
+
+
+def test_chat_json_falls_back_only_for_explicit_response_format_rejection():
+    unsupported = ProviderError(
+        status_code=400,
+        body={
+            "error": {
+                "param": "response_format",
+                "code": "unsupported_parameter",
+                "message": "response_format is not supported by this model",
+            }
+        },
+    )
+    sequence = CompletionSequence(unsupported, _response('{"ok": true}'))
+    client = _client_for(sequence)
+
+    result = client.chat_json(
+        messages=[{"role": "user", "content": "Return JSON"}],
+        max_attempts=1,
+    )
+
+    assert result == {"ok": True}
+    assert sequence.calls[0]["response_format"] == {"type": "json_object"}
+    assert "response_format" not in sequence.calls[1]
+
+
+def test_response_format_fallback_keeps_content_retry_available():
+    unsupported = ProviderError(
+        status_code=400,
+        body={
+            "error": {
+                "param": "response_format",
+                "code": "unsupported_parameter",
+                "message": "response_format is not supported by this model",
+            }
+        },
+    )
+    sequence = CompletionSequence(
+        unsupported,
+        _response('{"items": [', finish_reason="length"),
+        _response('{"items": [1]}'),
+    )
+    client = _client_for(sequence)
+
+    result = client.chat_json(
+        messages=[{"role": "user", "content": "Return JSON"}],
+        max_attempts=2,
+    )
+
+    assert result == {"items": [1]}
+    assert sequence.calls[0]["response_format"] == {"type": "json_object"}
+    assert "response_format" not in sequence.calls[1]
+    assert "response_format" not in sequence.calls[2]
+    assert "max_tokens" not in sequence.calls[2]
+
+
+def test_chat_json_defaults_to_one_content_attempt():
+    sequence = CompletionSequence(
+        _response(None),
+        _response('{"ok": true}'),
+    )
+    client = _client_for(sequence)
+
+    with pytest.raises(LLMResponseError, match="empty JSON"):
+        client.chat_json(
+            messages=[{"role": "user", "content": "Return JSON"}],
+        )
+
+    assert len(sequence.calls) == 1
+
+
+@pytest.mark.parametrize("status_code", [400, 401, 429, 500])
+def test_chat_json_does_not_retry_unrelated_provider_errors(status_code):
+    provider_error = ProviderError(
+        status_code=status_code,
+        body={
+            "error": {
+                "param": "messages",
+                "code": "invalid_request",
+                "message": "request failed for an unrelated reason",
+            }
+        },
+    )
+    sequence = CompletionSequence(provider_error, _response('{"ok": true}'))
+    client = _client_for(sequence)
+
+    with pytest.raises(ProviderError) as captured:
+        client.chat_json(
+            messages=[{"role": "user", "content": "Return JSON"}],
+            max_attempts=2,
+        )
+
+    assert captured.value is provider_error
+    assert len(sequence.calls) == 1
+
+
+def test_chat_json_reports_missing_choices_without_retrying_forever():
+    sequence = CompletionSequence(
+        _response(None, include_choice=False),
+        _response(None, include_choice=False),
+    )
+    client = _client_for(sequence)
+
+    with pytest.raises(LLMResponseError, match="no choices"):
+        client.chat_json(
+            messages=[{"role": "user", "content": "Return JSON"}],
+            max_attempts=2,
+        )
+
+    assert len(sequence.calls) == 2

backend/tests/test_ontology_api_errors.py

@@ -0,0 +1,70 @@
+import io
+
+from app import create_app
+from app.api import graph as graph_api
+from app.models.project import ProjectManager, ProjectStatus
+from app.utils.llm_client import LLMResponseError
+
+
+def _post_ontology(client):
+    return client.post(
+        "/api/graph/ontology/generate",
+        data={
+            "simulation_requirement": "Simulate the discussion.",
+            "files": (io.BytesIO(b"A short source document."), "source.md"),
+        },
+        content_type="multipart/form-data",
+    )
+
+
+def test_ontology_api_returns_safe_truncation_error_and_failed_project(
+    tmp_path,
+    monkeypatch,
+):
+    class FailingGenerator:
+        def generate(self, **kwargs):
+            raise LLMResponseError(
+                "LLM JSON output was truncated at the token limit",
+                finish_reason="length",
+            )
+
+    monkeypatch.setattr(ProjectManager, "PROJECTS_DIR", str(tmp_path))
+    monkeypatch.setattr(graph_api, "OntologyGenerator", FailingGenerator)
+
+    app = create_app()
+    app.config.update(TESTING=True)
+    response = _post_ontology(app.test_client())
+
+    assert response.status_code == 502
+    assert response.json["success"] is False
+    assert "token limit" in response.json["error"]
+    assert "traceback" not in response.json
+
+    project_id = response.json["data"]["project_id"]
+    project = ProjectManager.get_project(project_id)
+    assert project.status == ProjectStatus.FAILED
+    assert project.error == response.json["error"]
+
+
+def test_ontology_api_does_not_expose_provider_error_body(tmp_path, monkeypatch):
+    class ProviderError(RuntimeError):
+        status_code = 401
+        request_id = "request-safe-id"
+        body = {"error": {"message": "SECRET-PROVIDER-BODY"}}
+
+    class FailingGenerator:
+        def generate(self, **kwargs):
+            raise ProviderError("SECRET-PROVIDER-BODY")
+
+    monkeypatch.setattr(ProjectManager, "PROJECTS_DIR", str(tmp_path))
+    monkeypatch.setattr(graph_api, "OntologyGenerator", FailingGenerator)
+
+    app = create_app()
+    app.config.update(TESTING=True)
+    response = _post_ontology(app.test_client())
+
+    assert response.status_code == 502
+    assert "HTTP 401" in response.json["error"]
+    assert "request-safe-id" in response.json["error"]
+    assert "SECRET-PROVIDER-BODY" not in response.get_data(as_text=True)
+    assert "traceback" not in response.json

backend/tests/test_ontology_generator.py

@@ -1,6 +1,19 @@
 from app.services.ontology_generator import OntologyGenerator
 
 
+class RecordingLLMClient:
+    def __init__(self):
+        self.calls = []
+
+    def chat_json(self, **kwargs):
+        self.calls.append(kwargs)
+        return {
+            "entity_types": [],
+            "edge_types": [],
+            "analysis_summary": "ok",
+        }
+
+
 def _generator_for_test() -> OntologyGenerator:
     generator = OntologyGenerator(llm_client=object())
     generator.MAX_TEXT_LENGTH_FOR_LLM = 2000
@@ -50,3 +63,17 @@ def test_very_long_ontology_context_selects_representative_chunks():
     assert "BEGIN" in context
     assert "FINALEND" in context
     assert context.count("--- 文档 1 / 分块") == generator.MAX_LONG_TEXT_CHUNKS
+
+
+def test_ontology_generation_does_not_cap_structured_output_tokens():
+    llm = RecordingLLMClient()
+    generator = OntologyGenerator(llm_client=llm)
+
+    result = generator.generate(
+        document_texts=["A short source document."],
+        simulation_requirement="Simulate the public discussion.",
+    )
+
+    assert result["analysis_summary"] == "ok"
+    assert llm.calls[0]["max_tokens"] is None
+    assert llm.calls[0]["max_attempts"] == 2

frontend/src/api/index.js

@@ -37,6 +37,7 @@ service.interceptors.response.use(
   },
   error => {
     console.error('Response error:', error)
+    const apiError = error.response?.data?.error || error.response?.data?.message
     
     // 处理超时
     if (error.code === 'ECONNABORTED' && error.message.includes('timeout')) {
@@ -48,6 +49,12 @@ service.interceptors.response.use(
       console.error('Network error - please check your connection')
     }
 
+    // Axios rejects non-2xx responses before the success interceptor can
+    // surface the backend's safe, actionable error message.
+    if (typeof apiError === 'string' && apiError) {
+      error.message = apiError
+    }
+    
     return Promise.reject(error)
   }
 )