Errors don't contain logs for debugging - how to debug parsing errors?

[jmander11]

Describe the bug
When errors occur during searches there is no obvious path that can be taken to solve the issue.

For example, I see these logs when I make a search like "How do embedding models work?":
An error occurred while searching images: Error: Error parsing response from OpenAI: [
{
"expected": "object",
"code": "invalid_type",
"path": [],
"message": "Invalid input: expected object, received null"
}
]
An error occurred while searching videos: Error: Error parsing response from OpenAI: [
{
"expected": "object",
"code": "invalid_type",
"path": [],
"message": "Invalid input: expected object, received null"
}
]

What do I do with this information? I am running gpt-oss-20b and the response is not logged here for me to identify the issue. I'm going to have to go change vllm code to output the full llm generated message to understand what is happening here if there is no debug variable I can enable.

To Reproduce
Use gpt-oss-20b and make any search?

I have a 128k context window set and the tool parser/auto tool choice enabled with vllm.

Expected behavior
Ideally, I see the malformed snippet and entire response as 2 log messages when this occurs.

When the app tracebacks I can at least infer from the unhandled exception what the issue is, but I'm completely blind when it just says error parsing without telling me why :(

Additional context
I also couldn't get qwen3-30b-a3b to work, but that seems like it's because it's expected to use xml? It would probably be useful if the perplexica page mentions how to set up json parsing to work with perplexica and that a json model is necessary.

Is perplexica expecting the json in a separate tool call field or as part of the user message response? This would help me craft a user system message to possibly fix the malformed model output.

Thanks for any tips!

[ItzCrazyKns]

We're making use of the grammar enforced decoding (or structured outputs) that all major inference engines (and model providers) are providing. As soon as a query arrives, we use the LLM to classify it and that classification involves generation involving a strict JSON schema.

[ItzCrazyKns]

vLLM follows the OpenAI API format and has support for grammar enforced decoding so it should be generating proper JSON responses but its not so that's a bit weird.

[jmander11]

Is the reasoning text an issue or anything else here?

$ curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
"model": "/app/models/gpt-oss-20b-mxfp4",
"messages": [{"role": "user", "content": "Return a JSON with a key name and
value Bob"}],
"response_format": {"type": "json_object"}
}'
{"id":"chatcmpl-b3e31646d357b1b8","object":"chat.completion","created":1770790195,"model":"/app/models/gpt-oss-20b-mxfp4","choices":[{"index":0,"message":{"role":"assistant","content":"{"name":"Bob"}","refusal":null,"annotations":null,"audio":null,"function_call":null,"tool_calls":[],"reasoning":"User wants a JSON with key name and value Bob. So output JSON: {"name":"Bob"}. Should be in proper JSON. Provide that.","reasoning_content":"User wants a JSON with key name and value Bob. So output JSON: {"name":"Bob"}. Should be in proper JSON. Provide that."},"logprobs":null,"finish_reason":"stop","stop_reason":null,"token_ids":null}],"service_tier":null,"system_fingerprint":null,"usage":{"prompt_tokens":79,"total_tokens":124,"completion_tokens":45,"prompt_tokens_details":null},"prompt_logprobs":null,"prompt_token_ids":null,"kv_transfer_params":null}

[jmander11]

@ItzCrazyKns What does the json response need to look like? Are extra fields like reasoning/reasoning_content an issue?