BerriAI/litellm · error · DatabricksException

KeyError: , Got unexpected response from Databricks

Error message

KeyError: {e}, Got unexpected response from Databricks: {chunk}

What it means

The Databricks streaming chunk translator indexes chunk['id'], chunk['created'], chunk['model'] and choice['delta'] directly when building ModelResponseStream. A chunk missing any of those keys raises KeyError, re-raised as DatabricksException (status 400) naming the missing key and echoing the offending chunk. The stream contained a chunk whose shape the parser assumes but that chunk did not carry.

Solutions

1. Log the chunk from the message — it names the missing key and the exact payload; compare with the OpenAI streaming chunk schema
2. Ensure the endpoint is served OpenAI-compatible and reached via the databricks provider's expected invocations path
3. Retry the same call with stream=False — if it succeeds, the issue is chunk-shape-specific to streaming
4. Update LiteLLM (chunk translation is hardened frequently); if the payload matches Databricks docs, file an issue with the logged chunk

Example fix

# before
resp = litellm.completion(model='databricks/my-endpoint', messages=msgs, stream=True)  # KeyError chunk

# after — degrade to non-streaming when the stream schema breaks
try:
    resp = litellm.completion(model='databricks/my-endpoint', messages=msgs, stream=True)
except litellm.exceptions.APIError as e:
    if 'Got unexpected response from Databricks' in str(e):
        resp = litellm.completion(model='databricks/my-endpoint', messages=msgs, stream=False)
    else:
        raise

try:
    stream = litellm.completion(model='databricks/my-ep', messages=msgs, stream=True)
    for chunk in stream:
        handle(chunk)
except Exception as e:
    if 'Got unexpected response from Databricks' in str(e):
        resp = litellm.completion(model='databricks/my-ep', messages=msgs, stream=False)
        handle(resp)
    else:
        raise

When it happens

Trigger: Streaming from a Databricks serving endpoint whose chunks omit 'model' or 'created' (seen with some external-model gateways); an interleaved usage/keep-alive chunk without id; routing a non-OpenAI-compatible model through the databricks provider's streaming path.

Common situations: External/foundation model gateways emitting minimal chunk shapes; intermediary inserting heartbeat frames; serving API version differences dropping fields LiteLLM treats as required.

Related errors

• Failed to decode JSON from chunk
• Error receiving chunk from stream
• A2A send_message_streaming failed: no response received…
• AgentCore: Failed to read/parse JSON response body
• api_base is required for Pydantic AI agents

AI-assisted analysis of BerriAI/litellm@77b7c6c40c (2026-08-18). Data as JSON: /api/errors/8795da6dd6160c01. Report an issue: GitHub.

Appendix: source

Thrown at litellm/llms/databricks/chat/transformation.py:739

                (
                    reasoning_content,
                    thinking_blocks,
                ) = DatabricksConfig.extract_reasoning_content(choice["delta"].get("content"))

                choice["delta"]["content"] = content_str
                choice["delta"]["reasoning_content"] = reasoning_content
                choice["delta"]["thinking_blocks"] = thinking_blocks
                translated_choices.append(choice)
            return ModelResponseStream(
                id=chunk["id"],
                object="chat.completion.chunk",
                created=chunk["created"],
                model=chunk["model"],
                choices=translated_choices,
                usage=chunk.get("usage"),
            )
        except KeyError as e:
            raise DatabricksException(
                message=f"KeyError: {e}, Got unexpected response from Databricks: {chunk}",
                status_code=400,
            )
        except Exception as e:
            raise e

View on GitHub (pinned to 77b7c6c40c)