ResolverParsingError when using vLLM/lmdeploy as backend (Qwen3-32B-GPTQ)

[wangjiawen2013]

When using vllm or lmdeploy as the inference backend via the OpenAI-compatible API, langextract fails with a ResolverParsingError: Failed to parse JSON content: Expecting value: line 1 column 1 (char 0).

Interestingly, the exact same code and prompt work perfectly when using Ollama. It seems the JSON output from vLLM/lmdeploy is not being correctly captured or parsed by the Resolver, even when the model returns a valid JSON string (verified via manual requests calls).

File ~/programs/miniconda3/envs/agents/lib/python3.12/site-packages/langextract/resolver.py:271, in Resolver.resolve(self, input_text, suppress_parse_errors, **kwargs)
    267     logging.exception(
    268         "Failed to parse input_text: %s, error: %s", input_text, e
    269     )
    270     return []
--> 271   raise ResolverParsingError(str(e)) from e

ResolverParsingError: Failed to parse JSON content: Expecting value: line 1 column 1 (char 0)

Environment
Model: Qwen3-32B-GPTQ-Int4
Backend: vLLM / lmdeploy (OpenAI-compatible server)

Current Config:

config = lx.factory.ModelConfig(
    model_id="vllm:http://localhost:8000/v1",
    provider="VLLMLanguageModel", 
    provider_kwargs=dict(
        temperature=0.7,
        max_tokens=1024,
        # Server connection settings
        timeout=60.0,
    ),
)

[ywan614]

fence_output=False,
use_schema_constraints=False

add arguments above to avoid this issue

[wangjiawen2013]

I tried, but it still fails in most cases (the success rate is below 20%）.