500 error on /api/graph/ontology/generate even with tiny TXT and qwen-plus

[dkh345-sketch]

Hello,

I have issues with MiroFish that I cloned already but when uploading the pdf or md, it gives 500 error
What I did:

Cloned repo at current main.

Ran npm install, npm run setup:all, npm run dev.

Set .env to use LLM_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1, LLM_MODEL_NAME=qwen-plus.

Uploaded a tiny TXT file (~99 characters) as reality seed and clicked to start ontology generation.

What I expected:

Ontology generation to complete and move to environment setup.

What actually happened:

UI shows “Graph build complete”, then error; system dashboard shows “Request failed with status code 500”.

Backend log shows repeated POST /api/graph/ontology/generate HTTP/1.1" 500 immediately after “文本提取完成，共 99 字符” and “调用 LLM 生成本体定义…”.

Environment:

macOS version, Python version, Node version, LLM provider (DashScope qwen-plus; also tried local Ollama qwen/gemma with same 500).

Screenshots/logs:

Attach the console screenshot (like the one you sent) and copy‑paste a few lines of the traceback / log (but not your API keys).

[dosubot]

Hi @dkh345-sketch! I'm Dosu and I’m helping the MiroFish team.

This is a frequently reported issue — the 500 error occurs right after "调用 LLM 生成本体定义…" because chat_json() in llm_client.py unconditionally sends response_format={"type": "json_object"} to the LLM provider [1]. Even though DashScope/qwen-plus is the officially supported provider, many setups (including some DashScope configurations and all Ollama models) either reject this parameter or return malformed JSON, causing json.loads() to fail and bubble up as a 500 [2] [3].

Immediate workaround — add fallback retry logic to backend/app/utils/llm_client.py in the chat_json() method:

def chat_json(self, messages, temperature=0.3, max_tokens=4096):
    try:
        response = self.chat(messages=messages, temperature=temperature,
                             max_tokens=max_tokens,
                             response_format={"type": "json_object"})
    except Exception:
        # Retry without response_format for unsupported providers
        response = self.chat(messages=messages, temperature=temperature,
                             max_tokens=max_tokens)
    # ... rest of existing JSON cleaning logic

This is the fix from PR #186 [4]. After patching, clear cached bytecode and restart:

find backend -type d -name __pycache__ -exec rm -rf {} +
npm run dev

Other things to check:

• Make sure your LLM_BASE_URL ends exactly with /compatible-mode/v1 for DashScope [5]
• Verify your LLM_API_KEY is valid — the code doesn't validate it before making the LLM call [6]
• Check the full traceback in the 500 response body (visible in browser DevTools → Network tab → failed request → Response) — the endpoint includes it in the JSON error response [7]

Context: There are several open PRs addressing this systematically — PR #580 adds an LLM_JSON_MODE env variable to disable JSON mode entirely, and PR #532 hardens JSON parsing with repair logic — but none are merged yet. The manual patch above is the recommended fix for now.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.

[rqd6f4g6zn-bit]

A fix for this is up in #626 — adds a 5-stage robust LLM JSON parser (_parse_llm_json) that handles the trailing-prose pattern that's causing the strict json.loads to fail. 8/8 unit tests pass, including the exact position 10243 pattern. PR: #626