What causes "The response was blocked by VertexAI. {chunk}" in BerriAI/litellm?

Adjust safety_settings per litellm/Vertex docs: set BLOCK_ONLY_HIGH or BLOCK_NONE for the tripped categories (subject to Google allowing relaxation).

BerriAI/litellm · error · Exception

The response was blocked by VertexAI. {chunk}

Error message

The response was blocked by VertexAI. {chunk}

What it means

Vertex AI branch: when a stream chunk has no text (e.g. 'Part has no text') and the candidate's finish_reason is SAFETY, litellm raises this Exception embedding the raw proto chunk. It means Gemini/Vertex blocked the response with safety filters.

Source

Thrown at litellm/litellm_core_utils/streaming_handler.py:1320

                                        },
                                        "type": "function",
                                    }
                                ],
                            )
                            _streaming_response: Final = StreamingChoices(delta=_delta_obj)
                            _model_response: Final = ModelResponseStream()
                            _model_response.choices = [_streaming_response]
                            response_obj = {"original_chunk": _model_response}
                        else:
                            raise original_exception
                    if (
                        hasattr(chunk.candidates[0], "finish_reason")
                        and chunk.candidates[0].finish_reason.name != "FINISH_REASON_UNSPECIFIED"
                    ):  # every non-final chunk in vertex ai has this
                        self.received_finish_reason = map_finish_reason(chunk.candidates[0].finish_reason.name)
                except Exception:
                    if chunk.candidates[0].finish_reason.name == "SAFETY":
                        raise Exception(f"The response was blocked by VertexAI. {chunk}")
            else:
                completion_obj["content"] = str(chunk)
        elif self.custom_llm_provider == "petals":
            if self.completion_stream is None or len(self.completion_stream) == 0:
                if self.received_finish_reason is not None:
                    raise StopIteration
                else:
                    self.received_finish_reason = "stop"
            chunk_size = 30
            stream = cast(Any, self.completion_stream)
            new_chunk = stream[:chunk_size]
            completion_obj["content"] = new_chunk
            self.completion_stream = stream[chunk_size:]
        elif self.custom_llm_provider == "palm":
            # fake streaming
            response_obj = {}
            if self.completion_stream is None or len(self.completion_stream) == 0:
                if self.received_finish_reason is not None:

View on GitHub (pinned to 6c2dcb801b)

Solutions

Adjust safety_settings per litellm/Vertex docs: set BLOCK_ONLY_HIGH or BLOCK_NONE for the tripped categories (subject to Google allowing relaxation).
Rephrase the prompt to avoid triggering categories; split sensitive steps into multiple calls.
Use a model version with different safety posture (e.g. latest Gemini versions).
Catch this exception and surface a user-facing 'content blocked' message instead of retrying — retrying the same input usually fails again.

Example fix

# before
resp = litellm.completion(model="vertex_ai/gemini-1.5-pro", messages=msgs, stream=True)
# after
resp = litellm.completion(
    model="vertex_ai/gemini-1.5-pro", messages=msgs, stream=True,
    safety_settings=[
        {"category": "HARM_CATEGORY_HARASSMENT", "threshold": "BLOCK_ONLY_HIGH"},
        {"category": "HARM_CATEGORY_HATE_SPEECH", "threshold": "BLOCK_ONLY_HIGH"},
        {"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT", "threshold": "BLOCK_ONLY_HIGH"},
        {"category": "HARM_CATEGORY_DANGEROUS_CONTENT", "threshold": "BLOCK_ONLY_HIGH"},
    ],
)

Defensive patterns

Strategy: try-catch

Try / catch

try:
    for part in stream:
        ...
except Exception as e:
    if "blocked by VertexAI" in str(e):
        return ContentBlockedResponse(reason="safety_filter")  # do NOT retry same input
    raise

Prevention

Set explicit safety_settings thresholds on every Vertex/Gemini call.
Design UX around safety blocks instead of retry loops; the same input usually re-blocks.
Log blocked prompts (redacted) to identify which category trips.

When it happens

Trigger: Streaming a Vertex AI (Gemini) model where the prompt or partial generation trips safety filters; candidates carry finishReason=SAFETY and empty parts.

Common situations: Prompts touching violence/medical/self-harm topics; image+text prompts blocked by multimodal filters; default (unrelaxed) safety settings on newer Gemini models being stricter than expected.

Related errors

AI-assisted analysis of BerriAI/litellm@6c2dcb801b (2026-08-15). Data as JSON: /api/errors/b1841a9658d425d2. Report an issue: GitHub.