BerriAI/litellm · error · OpenAIError

{e} Original Response: {response.text}

Error message

{e}

Original Response: {response.text}

What it means

Error branch in streaming() (litellm/llms/openai/openai.py:1101) taken when the request failed but an httpx response with .text exists: litellm raises OpenAIError with message '<exception>\n\nOriginal Response: <response.text>' plus the upstream status, headers and body. The appended Original Response is the raw error body — typically a proxy's HTML error page or the provider's JSON error object — and is the primary diagnostic.

Source

Thrown at litellm/llms/openai/openai.py:1101

                ## check if body contains unprocessable params - related issue https://github.com/BerriAI/litellm/issues/4800
                if litellm.drop_params is True or drop_params is True:
                    data = drop_params_from_unprocessable_entity_error(e, data)
                else:
                    raise e
            except (
                Exception
            ) as e:  # need to exception handle here. async exceptions don't get caught in sync functions.
                if isinstance(e, OpenAIError):
                    raise e

                error_headers = getattr(e, "headers", None)
                status_code = getattr(e, "status_code", 500)
                error_response = getattr(e, "response", None)
                exception_body = getattr(e, "body", None)
                if error_headers is None and error_response:
                    error_headers = getattr(error_response, "headers", None)
                if response is not None and hasattr(response, "text"):
                    raise OpenAIError(
                        status_code=status_code,
                        message=f"{e}\n\nOriginal Response: {response.text}",
                        headers=error_headers,
                        body=exception_body,
                    )
                else:
                    if type(e).__name__ == "ReadTimeout":
                        raise OpenAIError(
                            status_code=408,
                            message=f"{type(e).__name__}",
                            headers=error_headers,
                            body=exception_body,
                        )
                    elif hasattr(e, "status_code"):
                        raise OpenAIError(
                            status_code=getattr(e, "status_code", 500),
                            message=str(e),
                            headers=error_headers,

View on GitHub (pinned to 77b7c6c40c)

Solutions

  1. Read the 'Original Response:' section — HTML means proxy/routing problem, JSON means provider error with a code
  2. Fix api_base so it resolves to the real /v1/chat/completions route
  3. If it is a 502/504 from a proxy, fix or bypass the proxy
  4. Map the status_code plus the JSON error code to the concrete fix (key, model, quota)

Example fix

# before
litellm.completion(model='gpt-4o', messages=msgs, stream=True, api_base='https://gw.internal')

# after
litellm.completion(model='gpt-4o', messages=msgs, stream=True, api_base='https://gw.internal/v1')
Defensive patterns

Strategy: try-catch

Validate before calling

import httpx

def stream_endpoint_ok(api_base: str, api_key: str) -> bool:
    r = httpx.post(
        f'{api_base.rstrip("/")}/chat/completions',
        headers={'Authorization': f'Bearer {api_key}'},
        json={'model': 'ping', 'messages': [{'role': 'user', 'content': 'ping'}], 'stream': True},
        timeout=15,
    )
    return r.status_code == 200 and 'text/event-stream' in r.headers.get('content-type', '')

Try / catch

from litellm.llms.openai.common_utils import OpenAIError

try:
    stream = litellm.completion(model=m, messages=msgs, stream=True)
    for chunk in stream:
        process(chunk)
except OpenAIError as e:
    if 'Original Response' in e.message:
        if '<html' in e.message.lower():
            alert_ops('proxy error page returned — check api_base path and upstream health')
        raise

Prevention

When it happens

Trigger: Streaming calls where the server answers non-200 with a body: nginx/traefik 502/504 HTML pages, OpenAI JSON errors (invalid_api_key, model_not_found), gateways returning HTML 404s because the api_base path is wrong, WAFs intercepting streaming POSTs.

Common situations: api_base missing /v1 so the path 404s with HTML; upstream LLM unavailable behind a load balancer; auth failures; corporate proxies blocking SSE.

Related errors

AI-assisted analysis of BerriAI/litellm@77b7c6c40c (2026-08-18). Data as JSON: /api/errors/841c359ab21a1649. Report an issue: GitHub.