BerriAI/litellm · error · ValueError

chunk

Error message

chunk

What it means

Second guard in the Predibase handler: if the raw (non-SSE) chunk string itself contains the substring 'error', the whole chunk is re-raised as ValueError, because a chunk-level 'error' token means the stream body is an error response rather than token data.

Source

Thrown at litellm/litellm_core_utils/streaming_handler.py:491

                print_verbose(f"data json: {data_json}")
                if "token" in data_json and "text" in data_json["token"]:
                    text = data_json["token"]["text"]
                if data_json.get("details", False) and data_json["details"].get("finish_reason", False):
                    is_finished = True
                    finish_reason = data_json["details"]["finish_reason"]
                elif data_json.get("generated_text", False):  # if full generated text exists, then stream is complete
                    text = ""  # don't return the final bos token
                    is_finished = True
                    finish_reason = "stop"
                elif data_json.get("error", False):
                    raise Exception(data_json.get("error"))
                return {
                    "text": text,
                    "is_finished": is_finished,
                    "finish_reason": finish_reason,
                }
            elif "error" in chunk:
                raise ValueError(chunk)
            return {
                "text": text,
                "is_finished": is_finished,
                "finish_reason": finish_reason,
            }
        except Exception as e:
            raise e

    def handle_ai21_chunk(self, chunk):  # fake streaming
        chunk = chunk.decode("utf-8")
        data_json: Final[_Ai21StreamData] = json.loads(chunk)
        try:
            text: Final = data_json["completions"][0]["data"]["text"]
            is_finished: Final = True
            finish_reason: Final = "stop"
            return {
                "text": text,
                "is_finished": is_finished,

View on GitHub (pinned to 6c2dcb801b)

Solutions

1. Print the chunk captured in the ValueError to identify the producer (Predibase vs proxy).
2. Fix the underlying request (auth, model id) that caused the error response.
3. If a proxy injects error pages, bypass it or fix its configuration for SSE endpoints.

try:
    for part in stream:
        ...
except ValueError as e:
    if "predibase" in str(e) or "error" in str(e).lower():
        diagnose_from_chunk(str(e))

When it happens

Trigger: Non-stream-shaped or error responses from Predibase that reach the streaming parser — e.g. a JSON error body delivered as one chunk, or proxy/gateway error text containing the word 'error'.

Common situations: Corporate proxies returning HTML/text error pages; Predibase returning a non-SSE JSON error; mismatch between stream=True and the endpoint's actual behavior.

Related errors

• data_json.get("error")
• An unknown error occurred with the stream
• Error receiving chunk from stream: {e}
• Braintrust API error: {e.response.text}
• Failed to connect to Braintrust API: {str(e)}

AI-assisted analysis of BerriAI/litellm@6c2dcb801b (2026-08-15). Data as JSON: /api/errors/fddb6e4642baccd1. Report an issue: GitHub.