converse_stream

Operation

converse_stream async

converse_stream(input: ConverseStreamInput, plugins: list[Plugin] | None = None) -> OutputEventStream[ConverseStreamOutput, ConverseStreamOperationOutput]

Sends messages to the specified Amazon Bedrock model and returns the response in a stream. ConverseStream provides a consistent API that works with all Amazon Bedrock models that support messages. This allows you to write code once and use it with different models. Should a model have unique inference parameters, you can also pass those unique parameters to the model.

To find out if a model supports streaming, call GetFoundationModel and check the responseStreamingSupported field in the response.

Note

The CLI doesn't support streaming operations in Amazon Bedrock, including ConverseStream.

Amazon Bedrock doesn't store any text, images, or documents that you provide as content. The data is only used to generate the response.

You can submit a prompt by including it in the messages field, specifying the modelId of a foundation model or inference profile to run inference on it, and including any other fields that are relevant to your use case.

You can also submit a prompt from Prompt management by specifying the ARN of the prompt version and including a map of variables to values in the promptVariables field. You can append more messages to the prompt by using the messages field. If you use a prompt from Prompt management, you can't include the following fields in the request: additionalModelRequestFields, inferenceConfig, system, or toolConfig. Instead, these fields must be defined through Prompt management. For more information, see Use a prompt from Prompt management.

For information about the Converse API, see Use the Converse API in the Amazon Bedrock User Guide. To use a guardrail, see Use a guardrail with the Converse API in the Amazon Bedrock User Guide. To use a tool with a model, see Tool use (Function calling) in the Amazon Bedrock User Guide

For example code, see Conversation streaming example in the Amazon Bedrock User Guide.

This operation requires permission for the bedrock:InvokeModelWithResponseStream action.

Warning

To deny all inference access to resources that you specify in the modelId field, you need to deny access to the bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream actions. Doing this also denies access to the resource through the base inference actions (InvokeModel and InvokeModelWithResponseStream). For more information see Deny access for inference on specific models.

For troubleshooting some of the common errors you might encounter when using the ConverseStream API, see Troubleshooting Amazon Bedrock API Error Codes in the Amazon Bedrock User Guide

Parameters:

Name	Type	Description	Default
input	ConverseStreamInput	An instance of ConverseStreamInput.	required
plugins	list[Plugin] | None	A list of callables that modify the configuration dynamically. Changes made by these plugins only apply for the duration of the operation execution and will not affect any other operation invocations.	None

Returns:

Type	Description
OutputEventStream[ConverseStreamOutput, ConverseStreamOperationOutput]	An OutputEventStream for server-to-client streaming of ConverseStreamOutput events with initial ConverseStreamOperationOutput response.

Source code in src/aws_sdk_bedrock_runtime/client.py

async def converse_stream(
    self, input: ConverseStreamInput, plugins: list[Plugin] | None = None
) -> OutputEventStream[ConverseStreamOutput, ConverseStreamOperationOutput]:
    """Sends messages to the specified Amazon Bedrock model and returns the
    response in a stream. `ConverseStream` provides a consistent API that
    works with all Amazon Bedrock models that support messages. This allows
    you to write code once and use it with different models. Should a model
    have unique inference parameters, you can also pass those unique
    parameters to the model.

    To find out if a model supports streaming, call
    [GetFoundationModel](https://docs.aws.amazon.com/bedrock/latest/APIReference/API_GetFoundationModel.html)
    and check the `responseStreamingSupported` field in the response.

    Note:
        The CLI doesn't support streaming operations in Amazon Bedrock,
        including `ConverseStream`.

    Amazon Bedrock doesn't store any text, images, or documents that you
    provide as content. The data is only used to generate the response.

    You can submit a prompt by including it in the `messages` field,
    specifying the `modelId` of a foundation model or inference profile to
    run inference on it, and including any other fields that are relevant to
    your use case.

    You can also submit a prompt from Prompt management by specifying the
    ARN of the prompt version and including a map of variables to values in
    the `promptVariables` field. You can append more messages to the prompt
    by using the `messages` field. If you use a prompt from Prompt
    management, you can't include the following fields in the request:
    `additionalModelRequestFields`, `inferenceConfig`, `system`, or
    `toolConfig`. Instead, these fields must be defined through Prompt
    management. For more information, see [Use a prompt from Prompt
    management](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-management-use.html).

    For information about the Converse API, see *Use the Converse API* in
    the *Amazon Bedrock User Guide*. To use a guardrail, see *Use a
    guardrail with the Converse API* in the *Amazon Bedrock User Guide*. To
    use a tool with a model, see *Tool use (Function calling)* in the
    *Amazon Bedrock User Guide*

    For example code, see *Conversation streaming example* in the *Amazon
    Bedrock User Guide*.

    This operation requires permission for the
    `bedrock:InvokeModelWithResponseStream` action.

    Warning:
        To deny all inference access to resources that you specify in the
        modelId field, you need to deny access to the `bedrock:InvokeModel` and
        `bedrock:InvokeModelWithResponseStream` actions. Doing this also denies
        access to the resource through the base inference actions
        ([InvokeModel](https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_InvokeModel.html)
        and
        [InvokeModelWithResponseStream](https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_InvokeModelWithResponseStream.html)).
        For more information see [Deny access for inference on specific
        models](https://docs.aws.amazon.com/bedrock/latest/userguide/security_iam_id-based-policy-examples.html#security_iam_id-based-policy-examples-deny-inference).

    For troubleshooting some of the common errors you might encounter when
    using the `ConverseStream` API, see [Troubleshooting Amazon Bedrock API
    Error
    Codes](https://docs.aws.amazon.com/bedrock/latest/userguide/troubleshooting-api-error-codes.html)
    in the Amazon Bedrock User Guide

    Args:
        input:
            An instance of `ConverseStreamInput`.
        plugins:
            A list of callables that modify the configuration dynamically.
            Changes made by these plugins only apply for the duration of the
            operation execution and will not affect any other operation
            invocations.

    Returns:
        An `OutputEventStream` for server-to-client streaming of
            `ConverseStreamOutput` events with initial `ConverseStreamOperationOutput`
            response.
    """
    operation_plugins: list[Plugin] = []
    if plugins:
        operation_plugins.extend(plugins)
    config = deepcopy(self._config)
    for plugin in operation_plugins:
        plugin(config)
    if config.protocol is None or config.transport is None:
        raise ExpectationNotMetError(
            "protocol and transport MUST be set on the config to make calls."
        )
    pipeline = RequestPipeline(protocol=config.protocol, transport=config.transport)
    call = ClientCall(
        input=input,
        operation=CONVERSE_STREAM,
        context=TypedProperties({"config": config}),
        interceptor=InterceptorChain(config.interceptors),
        auth_scheme_resolver=config.auth_scheme_resolver,
        supported_auth_schemes=config.auth_schemes,
        endpoint_resolver=config.endpoint_resolver,
        retry_strategy=config.retry_strategy,
    )

    return await pipeline.output_stream(
        call, ConverseStreamOutput, _ConverseStreamOutputDeserializer().deserialize
    )

Input

ConverseStreamInput dataclass

Dataclass for ConverseStreamInput structure.

Source code in src/aws_sdk_bedrock_runtime/models.py

@dataclass(kw_only=True)
class ConverseStreamInput:
    """Dataclass for ConverseStreamInput structure."""

    model_id: str | None = None
    """Specifies the model or throughput with which to run inference, or the
    prompt resource to use in inference. The value depends on the resource
    that you use:

    - If you use a base model, specify the model ID or its ARN. For a list
      of model IDs for base models, see [Amazon Bedrock base model IDs
      (on-demand
      throughput)](https://docs.aws.amazon.com/bedrock/latest/userguide/model-ids.html#model-ids-arns)
      in the Amazon Bedrock User Guide.

    - If you use an inference profile, specify the inference profile ID or
      its ARN. For a list of inference profile IDs, see [Supported Regions
      and models for cross-region
      inference](https://docs.aws.amazon.com/bedrock/latest/userguide/cross-region-inference-support.html)
      in the Amazon Bedrock User Guide.

    - If you use a provisioned model, specify the ARN of the Provisioned
      Throughput. For more information, see [Run inference using a
      Provisioned
      Throughput](https://docs.aws.amazon.com/bedrock/latest/userguide/prov-thru-use.html)
      in the Amazon Bedrock User Guide.

    - If you use a custom model, first purchase Provisioned Throughput for
      it. Then specify the ARN of the resulting provisioned model. For more
      information, see [Use a custom model in Amazon
      Bedrock](https://docs.aws.amazon.com/bedrock/latest/userguide/model-customization-use.html)
      in the Amazon Bedrock User Guide.

    - To include a prompt that was defined in [Prompt
      management](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-management.html),
      specify the ARN of the prompt version to use.

    The Converse API doesn't support [imported
    models](https://docs.aws.amazon.com/bedrock/latest/userguide/model-customization-import-model.html).
    """

    messages: list[Message] | None = None
    """The messages that you want to send to the model."""

    system: list[SystemContentBlock] | None = None
    """A prompt that provides instructions or context to the model about the
    task it should perform, or the persona it should adopt during the
    conversation.
    """

    inference_config: InferenceConfiguration | None = None
    """Inference parameters to pass to the model. `Converse` and
    `ConverseStream` support a base set of inference parameters. If you need
    to pass additional parameters that the model supports, use the
    `additionalModelRequestFields` request field.
    """

    tool_config: ToolConfiguration | None = None
    """Configuration information for the tools that the model can use when
    generating a response.

    For information about models that support streaming tool use, see
    [Supported models and model
    features](https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html#conversation-inference-supported-models-features).
    """

    guardrail_config: GuardrailStreamConfiguration | None = None
    """Configuration information for a guardrail that you want to use in the
    request. If you include `guardContent` blocks in the `content` field in
    the `messages` field, the guardrail operates only on those messages. If
    you include no `guardContent` blocks, the guardrail operates on all
    messages in the request body and in any included prompt resource.
    """

    additional_model_request_fields: Document | None = None
    """Additional inference parameters that the model supports, beyond the base
    set of inference parameters that `Converse` and `ConverseStream` support
    in the `inferenceConfig` field. For more information, see [Model
    parameters](https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters.html).
    """

    prompt_variables: dict[str, PromptVariableValues] | None = field(
        repr=False, default=None
    )
    """Contains a map of variables in a prompt from Prompt management to
    objects containing the values to fill in for them when running model
    invocation. This field is ignored if you don't specify a prompt
    resource in the `modelId` field.
    """

    additional_model_response_field_paths: list[str] | None = None
    """Additional model parameters field paths to return in the response.
    `Converse` and `ConverseStream` return the requested fields as a JSON
    Pointer object in the `additionalModelResponseFields` field. The
    following is example JSON for `additionalModelResponseFieldPaths`.

    `[ "/stop_sequence" ]`

    For information about the JSON Pointer syntax, see the [Internet
    Engineering Task Force
    (IETF)](https://datatracker.ietf.org/doc/html/rfc6901) documentation.

    `Converse` and `ConverseStream` reject an empty JSON Pointer or
    incorrectly structured JSON Pointer with a `400` error code. if the JSON
    Pointer is valid, but the requested field is not in the model response,
    it is ignored by `Converse`.
    """

    request_metadata: dict[str, str] | None = field(repr=False, default=None)
    """Key-value pairs that you can use to filter invocation logs."""

    performance_config: PerformanceConfiguration | None = None
    """Model performance settings for the request."""

    def serialize(self, serializer: ShapeSerializer):
        serializer.write_struct(_SCHEMA_CONVERSE_STREAM_INPUT, self)

    def serialize_members(self, serializer: ShapeSerializer):
        if self.model_id is not None:
            serializer.write_string(
                _SCHEMA_CONVERSE_STREAM_INPUT.members["modelId"], self.model_id
            )

        if self.messages is not None:
            _serialize_messages(
                serializer,
                _SCHEMA_CONVERSE_STREAM_INPUT.members["messages"],
                self.messages,
            )

        if self.system is not None:
            _serialize_system_content_blocks(
                serializer, _SCHEMA_CONVERSE_STREAM_INPUT.members["system"], self.system
            )

        if self.inference_config is not None:
            serializer.write_struct(
                _SCHEMA_CONVERSE_STREAM_INPUT.members["inferenceConfig"],
                self.inference_config,
            )

        if self.tool_config is not None:
            serializer.write_struct(
                _SCHEMA_CONVERSE_STREAM_INPUT.members["toolConfig"], self.tool_config
            )

        if self.guardrail_config is not None:
            serializer.write_struct(
                _SCHEMA_CONVERSE_STREAM_INPUT.members["guardrailConfig"],
                self.guardrail_config,
            )

        if self.additional_model_request_fields is not None:
            serializer.write_document(
                _SCHEMA_CONVERSE_STREAM_INPUT.members["additionalModelRequestFields"],
                self.additional_model_request_fields,
            )

        if self.prompt_variables is not None:
            _serialize_prompt_variable_map(
                serializer,
                _SCHEMA_CONVERSE_STREAM_INPUT.members["promptVariables"],
                self.prompt_variables,
            )

        if self.additional_model_response_field_paths is not None:
            _serialize_additional_model_response_field_paths(
                serializer,
                _SCHEMA_CONVERSE_STREAM_INPUT.members[
                    "additionalModelResponseFieldPaths"
                ],
                self.additional_model_response_field_paths,
            )

        if self.request_metadata is not None:
            _serialize_request_metadata(
                serializer,
                _SCHEMA_CONVERSE_STREAM_INPUT.members["requestMetadata"],
                self.request_metadata,
            )

        if self.performance_config is not None:
            serializer.write_struct(
                _SCHEMA_CONVERSE_STREAM_INPUT.members["performanceConfig"],
                self.performance_config,
            )

    @classmethod
    def deserialize(cls, deserializer: ShapeDeserializer) -> Self:
        return cls(**cls.deserialize_kwargs(deserializer))

    @classmethod
    def deserialize_kwargs(cls, deserializer: ShapeDeserializer) -> dict[str, Any]:
        kwargs: dict[str, Any] = {}

        def _consumer(schema: Schema, de: ShapeDeserializer) -> None:
            match schema.expect_member_index():
                case 0:
                    kwargs["model_id"] = de.read_string(
                        _SCHEMA_CONVERSE_STREAM_INPUT.members["modelId"]
                    )

                case 1:
                    kwargs["messages"] = _deserialize_messages(
                        de, _SCHEMA_CONVERSE_STREAM_INPUT.members["messages"]
                    )

                case 2:
                    kwargs["system"] = _deserialize_system_content_blocks(
                        de, _SCHEMA_CONVERSE_STREAM_INPUT.members["system"]
                    )

                case 3:
                    kwargs["inference_config"] = InferenceConfiguration.deserialize(de)

                case 4:
                    kwargs["tool_config"] = ToolConfiguration.deserialize(de)

                case 5:
                    kwargs["guardrail_config"] = (
                        GuardrailStreamConfiguration.deserialize(de)
                    )

                case 6:
                    kwargs["additional_model_request_fields"] = de.read_document(
                        _SCHEMA_CONVERSE_STREAM_INPUT.members[
                            "additionalModelRequestFields"
                        ]
                    )

                case 7:
                    kwargs["prompt_variables"] = _deserialize_prompt_variable_map(
                        de, _SCHEMA_CONVERSE_STREAM_INPUT.members["promptVariables"]
                    )

                case 8:
                    kwargs["additional_model_response_field_paths"] = (
                        _deserialize_additional_model_response_field_paths(
                            de,
                            _SCHEMA_CONVERSE_STREAM_INPUT.members[
                                "additionalModelResponseFieldPaths"
                            ],
                        )
                    )

                case 9:
                    kwargs["request_metadata"] = _deserialize_request_metadata(
                        de, _SCHEMA_CONVERSE_STREAM_INPUT.members["requestMetadata"]
                    )

                case 10:
                    kwargs["performance_config"] = PerformanceConfiguration.deserialize(
                        de
                    )

                case _:
                    logger.debug("Unexpected member schema: %s", schema)

        deserializer.read_struct(_SCHEMA_CONVERSE_STREAM_INPUT, consumer=_consumer)
        return kwargs

Attributes

additional_model_request_fields class-attribute instance-attribute

additional_model_request_fields: Document | None = None

Additional inference parameters that the model supports, beyond the base set of inference parameters that Converse and ConverseStream support in the inferenceConfig field. For more information, see Model parameters.

additional_model_response_field_paths class-attribute instance-attribute

additional_model_response_field_paths: list[str] | None = None

Additional model parameters field paths to return in the response. Converse and ConverseStream return the requested fields as a JSON Pointer object in the additionalModelResponseFields field. The following is example JSON for additionalModelResponseFieldPaths.

[ "/stop_sequence" ]

For information about the JSON Pointer syntax, see the Internet Engineering Task Force (IETF) documentation.

Converse and ConverseStream reject an empty JSON Pointer or incorrectly structured JSON Pointer with a 400 error code. if the JSON Pointer is valid, but the requested field is not in the model response, it is ignored by Converse.

guardrail_config class-attribute instance-attribute

guardrail_config: GuardrailStreamConfiguration | None = None

Configuration information for a guardrail that you want to use in the request. If you include guardContent blocks in the content field in the messages field, the guardrail operates only on those messages. If you include no guardContent blocks, the guardrail operates on all messages in the request body and in any included prompt resource.

inference_config class-attribute instance-attribute

inference_config: InferenceConfiguration | None = None

Inference parameters to pass to the model. Converse and ConverseStream support a base set of inference parameters. If you need to pass additional parameters that the model supports, use the additionalModelRequestFields request field.

messages class-attribute instance-attribute

messages: list[Message] | None = None

The messages that you want to send to the model.

model_id class-attribute instance-attribute

model_id: str | None = None

Specifies the model or throughput with which to run inference, or the prompt resource to use in inference. The value depends on the resource that you use:

• If you use a base model, specify the model ID or its ARN. For a list of model IDs for base models, see Amazon Bedrock base model IDs (on-demand throughput) in the Amazon Bedrock User Guide.

• If you use an inference profile, specify the inference profile ID or its ARN. For a list of inference profile IDs, see Supported Regions and models for cross-region inference in the Amazon Bedrock User Guide.

• If you use a provisioned model, specify the ARN of the Provisioned Throughput. For more information, see Run inference using a Provisioned Throughput in the Amazon Bedrock User Guide.

• If you use a custom model, first purchase Provisioned Throughput for it. Then specify the ARN of the resulting provisioned model. For more information, see Use a custom model in Amazon Bedrock in the Amazon Bedrock User Guide.

• To include a prompt that was defined in Prompt management, specify the ARN of the prompt version to use.

The Converse API doesn't support imported models.

performance_config class-attribute instance-attribute

performance_config: PerformanceConfiguration | None = None

Model performance settings for the request.

prompt_variables class-attribute instance-attribute

prompt_variables: dict[str, PromptVariableValues] | None = field(repr=False, default=None)

Contains a map of variables in a prompt from Prompt management to objects containing the values to fill in for them when running model invocation. This field is ignored if you don't specify a prompt resource in the modelId field.

request_metadata class-attribute instance-attribute

request_metadata: dict[str, str] | None = field(repr=False, default=None)

Key-value pairs that you can use to filter invocation logs.

system class-attribute instance-attribute

system: list[SystemContentBlock] | None = None

A prompt that provides instructions or context to the model about the task it should perform, or the persona it should adopt during the conversation.

tool_config class-attribute instance-attribute

tool_config: ToolConfiguration | None = None

Configuration information for the tools that the model can use when generating a response.

For information about models that support streaming tool use, see Supported models and model features.

Output

This operation returns an OutputEventStream for server-to-client streaming.

Event Stream Structure

Output Event Type

ConverseStreamOutput

Initial Response Structure

ConverseStreamOperationOutput dataclass

Dataclass for ConverseStreamOperationOutput structure.

Source code in src/aws_sdk_bedrock_runtime/models.py

@dataclass(kw_only=True)
class ConverseStreamOperationOutput:
    """Dataclass for ConverseStreamOperationOutput structure."""

    def serialize(self, serializer: ShapeSerializer):
        serializer.write_struct(_SCHEMA_CONVERSE_STREAM_OPERATION_OUTPUT, self)

    def serialize_members(self, serializer: ShapeSerializer):
        pass

    @classmethod
    def deserialize(cls, deserializer: ShapeDeserializer) -> Self:
        return cls(**cls.deserialize_kwargs(deserializer))

    @classmethod
    def deserialize_kwargs(cls, deserializer: ShapeDeserializer) -> dict[str, Any]:
        kwargs: dict[str, Any] = {}

        def _consumer(schema: Schema, de: ShapeDeserializer) -> None:
            match schema.expect_member_index():
                case _:
                    logger.debug("Unexpected member schema: %s", schema)

        deserializer.read_struct(
            _SCHEMA_CONVERSE_STREAM_OPERATION_OUTPUT, consumer=_consumer
        )
        return kwargs

Errors

• AccessDeniedException

• InternalServerException

• ModelErrorException

• ModelNotReadyException

• ModelTimeoutException

• ResourceNotFoundException

• ServiceUnavailableException

• ThrottlingException

• ValidationException