mirascope 1.25.7__py3-none-any.whl → 2.0.0a0__py3-none-any.whl

mirascope/llm/clients/anthropic/_utils/decode.py

@@ -0,0 +1,244 @@
+"""Anthropic response decoding."""
+
+import json
+from typing import Any, TypeAlias, cast
+
+from anthropic import types as anthropic_types
+from anthropic.lib.streaming import AsyncMessageStreamManager, MessageStreamManager
+
+from ....content import (
+    AssistantContentPart,
+    Text,
+    TextChunk,
+    TextEndChunk,
+    TextStartChunk,
+    Thought,
+    ThoughtChunk,
+    ThoughtEndChunk,
+    ThoughtStartChunk,
+    ToolCall,
+    ToolCallChunk,
+    ToolCallEndChunk,
+    ToolCallStartChunk,
+)
+from ....messages import AssistantMessage
+from ....responses import (
+    AsyncChunkIterator,
+    ChunkIterator,
+    FinishReason,
+    FinishReasonChunk,
+    RawMessageChunk,
+    RawStreamEventChunk,
+)
+from ..model_ids import AnthropicModelId
+
+ANTHROPIC_FINISH_REASON_MAP = {
+    "max_tokens": FinishReason.MAX_TOKENS,
+    "refusal": FinishReason.REFUSAL,
+}
+
+
+def _decode_assistant_content(
+    content: anthropic_types.ContentBlock,
+) -> AssistantContentPart:
+    """Convert Anthropic content block to mirascope AssistantContentPart."""
+    if content.type == "text":
+        return Text(text=content.text)
+    elif content.type == "tool_use":
+        return ToolCall(
+            id=content.id,
+            name=content.name,
+            args=json.dumps(content.input),
+        )
+    elif content.type == "thinking":
+        return Thought(thought=content.thinking)
+    else:
+        raise NotImplementedError(
+            f"Support for content type `{content.type}` is not yet implemented."
+        )
+
+
+def decode_response(
+    response: anthropic_types.Message,
+    model_id: AnthropicModelId,
+) -> tuple[AssistantMessage, FinishReason | None]:
+    """Convert Anthropic message to mirascope AssistantMessage."""
+    assistant_message = AssistantMessage(
+        content=[_decode_assistant_content(part) for part in response.content],
+        provider="anthropic",
+        model_id=model_id,
+        raw_message={
+            "role": response.role,
+            "content": [part.model_dump() for part in response.content],
+        },
+    )
+    finish_reason = (
+        ANTHROPIC_FINISH_REASON_MAP.get(response.stop_reason)
+        if response.stop_reason
+        else None
+    )
+    return assistant_message, finish_reason
+
+
+ContentBlock: TypeAlias = (
+    anthropic_types.TextBlockParam
+    | anthropic_types.ThinkingBlockParam
+    | anthropic_types.ToolUseBlockParam
+    | anthropic_types.ThinkingBlockParam
+    | anthropic_types.RedactedThinkingBlockParam
+)
+
+
+class _AnthropicChunkProcessor:
+    """Processes Anthropic stream events and maintains state across events."""
+
+    def __init__(self) -> None:
+        self.current_block_param: ContentBlock | None = None
+        self.accumulated_tool_json: str = ""
+        self.accumulated_blocks: list[ContentBlock] = []
+
+    def process_event(
+        self, event: anthropic_types.RawMessageStreamEvent
+    ) -> ChunkIterator:
+        """Process a single Anthropic event and yield the appropriate content chunks."""
+        yield RawStreamEventChunk(raw_stream_event=event)
+
+        if event.type == "content_block_start":
+            content_block = event.content_block
+
+            if content_block.type == "text":
+                self.current_block_param = {
+                    "type": "text",
+                    "text": content_block.text,
+                }
+                yield TextStartChunk()
+            elif content_block.type == "tool_use":
+                self.current_block_param = {
+                    "type": "tool_use",
+                    "id": content_block.id,
+                    "name": content_block.name,
+                    "input": {},
+                }
+                self.accumulated_tool_json = ""
+                yield ToolCallStartChunk(
+                    id=content_block.id,
+                    name=content_block.name,
+                )
+            elif content_block.type == "thinking":
+                self.current_block_param = {
+                    "type": "thinking",
+                    "thinking": "",
+                    "signature": "",
+                }
+                yield ThoughtStartChunk()
+            elif content_block.type == "redacted_thinking":  # pragma: no cover
+                self.current_block_param = {
+                    "type": "redacted_thinking",
+                    "data": content_block.data,
+                }
+            else:
+                raise NotImplementedError
+
+        elif event.type == "content_block_delta":
+            if self.current_block_param is None:  # pragma: no cover
+                raise RuntimeError("Received delta without a current block")
+
+            delta = event.delta
+            if delta.type == "text_delta":
+                if self.current_block_param["type"] != "text":  # pragma: no cover
+                    raise RuntimeError(
+                        f"Received text_delta for {self.current_block_param['type']} block"
+                    )
+                self.current_block_param["text"] += delta.text
+                yield TextChunk(delta=delta.text)
+            elif delta.type == "input_json_delta":
+                if self.current_block_param["type"] != "tool_use":  # pragma: no cover
+                    raise RuntimeError(
+                        f"Received input_json_delta for {self.current_block_param['type']} block"
+                    )
+                self.accumulated_tool_json += delta.partial_json
+                yield ToolCallChunk(delta=delta.partial_json)
+            elif delta.type == "thinking_delta":
+                if self.current_block_param["type"] != "thinking":  # pragma: no cover
+                    raise RuntimeError(
+                        f"Received thinking_delta for {self.current_block_param['type']} block"
+                    )
+                self.current_block_param["thinking"] += delta.thinking
+                yield ThoughtChunk(delta=delta.thinking)
+            elif delta.type == "signature_delta":
+                if self.current_block_param["type"] != "thinking":  # pragma: no cover
+                    raise RuntimeError(
+                        f"Received signature_delta for {self.current_block_param['type']} block"
+                    )
+                self.current_block_param["signature"] += delta.signature
+            else:
+                raise RuntimeError(
+                    f"Received unsupported delta type: {delta.type}"
+                )  # pragma: no cover
+
+        elif event.type == "content_block_stop":
+            if self.current_block_param is None:  # pragma: no cover
+                raise RuntimeError("Received stop without a current block")
+
+            block_type = self.current_block_param["type"]
+
+            if block_type == "text":
+                yield TextEndChunk()
+            elif block_type == "tool_use":
+                if self.current_block_param["type"] != "tool_use":  # pragma: no cover
+                    raise RuntimeError(
+                        f"Block type mismatch: stored {self.current_block_param['type']}, expected tool_use"
+                    )
+                self.current_block_param["input"] = (
+                    json.loads(self.accumulated_tool_json)
+                    if self.accumulated_tool_json
+                    else {}
+                )
+                yield ToolCallEndChunk()
+            elif block_type == "thinking":
+                yield ThoughtEndChunk()
+            else:
+                raise NotImplementedError
+
+            self.accumulated_blocks.append(self.current_block_param)
+            self.current_block_param = None
+
+        elif event.type == "message_delta":
+            if event.delta.stop_reason:
+                finish_reason = ANTHROPIC_FINISH_REASON_MAP.get(event.delta.stop_reason)
+                if finish_reason is not None:
+                    yield FinishReasonChunk(finish_reason=finish_reason)
+
+    def raw_message_chunk(self) -> RawMessageChunk:
+        return RawMessageChunk(
+            raw_message=cast(
+                dict[str, Any],
+                {
+                    "role": "assistant",
+                    "content": self.accumulated_blocks,
+                },
+            )
+        )
+
+
+def decode_stream(
+    anthropic_stream_manager: MessageStreamManager,
+) -> ChunkIterator:
+    """Returns a ChunkIterator converted from an Anthropic MessageStreamManager"""
+    processor = _AnthropicChunkProcessor()
+    with anthropic_stream_manager as stream:
+        for event in stream._raw_stream:
+            yield from processor.process_event(event)
+    yield processor.raw_message_chunk()
+
+
+async def decode_async_stream(
+    anthropic_stream_manager: AsyncMessageStreamManager,
+) -> AsyncChunkIterator:
+    """Returns an AsyncChunkIterator converted from an Anthropic MessageStreamManager"""
+    processor = _AnthropicChunkProcessor()
+    async with anthropic_stream_manager as stream:
+        async for event in stream._raw_stream:
+            for item in processor.process_event(event):
+                yield item
+    yield processor.raw_message_chunk()

mirascope/llm/clients/anthropic/_utils/encode.py

@@ -0,0 +1,243 @@
+"""Anthropic message encoding and request preparation."""
+
+import json
+from collections.abc import Sequence
+from functools import lru_cache
+from typing import Literal, TypedDict, cast
+from typing_extensions import Required
+
+from anthropic import Omit, types as anthropic_types
+
+from ....content import ContentPart, ImageMimeType
+from ....exceptions import FeatureNotSupportedError, FormattingModeNotSupportedError
+from ....formatting import (
+    Format,
+    FormattableT,
+    _utils as _formatting_utils,
+    resolve_format,
+)
+from ....messages import AssistantMessage, Message, UserMessage
+from ....tools import FORMAT_TOOL_NAME, BaseToolkit, ToolSchema
+from ...base import Params, _utils as _base_utils
+from ..model_ids import AnthropicModelId
+
+DEFAULT_MAX_TOKENS = 16000
+
+AnthropicImageMimeType = Literal["image/jpeg", "image/png", "image/gif", "image/webp"]
+
+
+def encode_image_mime_type(
+    mime_type: ImageMimeType,
+) -> AnthropicImageMimeType:
+    """Convert an ImageMimeType into anthropic supported mime type"""
+    if mime_type in ("image/jpeg", "image/png", "image/gif", "image/webp"):
+        return mime_type
+    raise FeatureNotSupportedError(
+        feature=f"Image with mime_type: {mime_type}", provider="anthropic"
+    )  # pragma: no cover
+
+
+class MessageCreateKwargs(TypedDict, total=False):
+    """Kwargs for Anthropic Message.create method."""
+
+    model: Required[str]
+    max_tokens: Required[int]
+    messages: Sequence[anthropic_types.MessageParam]
+    system: str | Omit
+    tools: Sequence[anthropic_types.ToolParam] | Omit
+    tool_choice: anthropic_types.ToolChoiceParam | Omit
+    temperature: float | Omit
+    top_p: float | Omit
+    top_k: int | Omit
+    stop_sequences: list[str] | Omit
+    thinking: anthropic_types.ThinkingConfigParam | Omit
+
+
+def _encode_content(
+    content: Sequence[ContentPart], encode_thoughts: bool
+) -> str | Sequence[anthropic_types.ContentBlockParam]:
+    """Convert mirascope content to Anthropic content format."""
+
+    if len(content) == 1 and content[0].type == "text":
+        return content[0].text
+
+    blocks: list[anthropic_types.ContentBlockParam] = []
+
+    for part in content:
+        if part.type == "text":
+            blocks.append(anthropic_types.TextBlockParam(type="text", text=part.text))
+        elif part.type == "image":
+            source: (
+                anthropic_types.Base64ImageSourceParam
+                | anthropic_types.URLImageSourceParam
+            )
+            if part.source.type == "base64_image_source":
+                source = anthropic_types.Base64ImageSourceParam(
+                    type="base64",
+                    media_type=encode_image_mime_type(part.source.mime_type),
+                    data=part.source.data,
+                )
+            else:  # url_image_source
+                source = anthropic_types.URLImageSourceParam(
+                    type="url",
+                    url=part.source.url,
+                )
+            blocks.append(anthropic_types.ImageBlockParam(type="image", source=source))
+        elif part.type == "audio":
+            raise FeatureNotSupportedError(
+                "audio input",
+                "anthropic",
+                message="Anthropic does not support audio inputs.",
+            )
+        elif part.type == "tool_output":
+            blocks.append(
+                anthropic_types.ToolResultBlockParam(
+                    type="tool_result",
+                    tool_use_id=part.id,
+                    content=str(part.value),
+                )
+            )
+        elif part.type == "tool_call":
+            blocks.append(
+                anthropic_types.ToolUseBlockParam(
+                    type="tool_use",
+                    id=part.id,
+                    name=part.name,
+                    input=json.loads(part.args),
+                )
+            )
+        elif part.type == "thought":
+            if encode_thoughts:
+                blocks.append(
+                    anthropic_types.TextBlockParam(
+                        type="text", text="**Thinking:** " + part.thought
+                    )
+                )
+        else:
+            raise NotImplementedError(f"Unsupported content type: {part.type}")
+
+    return blocks
+
+
+def _encode_message(
+    message: UserMessage | AssistantMessage,
+    model_id: AnthropicModelId,
+    encode_thoughts: bool,
+) -> anthropic_types.MessageParam:
+    """Convert user or assistant `Message`s to Anthropic `MessageParam` format.
+
+    Args:
+        messages: A Sequence containing `UserMessage`s or `AssistantMessage`s
+        model_id: The Anthropic model ID being used
+
+    Returns:
+        A Sequence of converted Anthropic `MessageParam`
+    """
+
+    if (
+        message.role == "assistant"
+        and message.provider == "anthropic"
+        and message.model_id == model_id
+        and message.raw_message
+        and not encode_thoughts
+    ):
+        return cast(anthropic_types.MessageParam, message.raw_message)
+    return {
+        "role": message.role,
+        "content": _encode_content(message.content, encode_thoughts),
+    }
+
+
+@lru_cache(maxsize=128)
+def _convert_tool_to_tool_param(tool: ToolSchema) -> anthropic_types.ToolParam:
+    """Convert a single Mirascope tool to Anthropic tool format with caching."""
+    schema_dict = tool.parameters.model_dump(by_alias=True, exclude_none=True)
+    schema_dict["type"] = "object"
+    return anthropic_types.ToolParam(
+        name=tool.name,
+        description=tool.description,
+        input_schema=schema_dict,
+    )
+
+
+def encode_request(
+    *,
+    model_id: AnthropicModelId,
+    messages: Sequence[Message],
+    tools: Sequence[ToolSchema] | BaseToolkit | None,
+    format: type[FormattableT] | Format[FormattableT] | None,
+    params: Params,
+) -> tuple[Sequence[Message], Format[FormattableT] | None, MessageCreateKwargs]:
+    """Prepares a request for the `Anthropic.messages.create` method."""
+    kwargs: MessageCreateKwargs = MessageCreateKwargs(
+        {
+            "model": model_id,
+            "max_tokens": DEFAULT_MAX_TOKENS,
+        }
+    )
+    encode_thoughts = False
+
+    with _base_utils.ensure_all_params_accessed(
+        params=params, provider="anthropic", unsupported_params=["seed"]
+    ) as param_accessor:
+        if param_accessor.temperature is not None:
+            kwargs["temperature"] = param_accessor.temperature
+        if param_accessor.max_tokens is not None:
+            kwargs["max_tokens"] = param_accessor.max_tokens
+        if param_accessor.top_p is not None:
+            kwargs["top_p"] = param_accessor.top_p
+        if param_accessor.top_k is not None:
+            kwargs["top_k"] = param_accessor.top_k
+        if param_accessor.stop_sequences is not None:
+            kwargs["stop_sequences"] = param_accessor.stop_sequences
+        if param_accessor.thinking is not None:
+            if param_accessor.thinking:
+                # Set budget to 50% of max_tokens with minimum of 1024
+                budget_tokens = max(1024, kwargs["max_tokens"] // 2)
+                kwargs["thinking"] = {"type": "enabled", "budget_tokens": budget_tokens}
+            else:
+                kwargs["thinking"] = {"type": "disabled"}
+        if param_accessor.encode_thoughts_as_text:
+            encode_thoughts = True
+
+    tools = tools.tools if isinstance(tools, BaseToolkit) else tools or []
+    anthropic_tools = [_convert_tool_to_tool_param(tool) for tool in tools]
+    format = resolve_format(format, default_mode="tool")
+    if format is not None:
+        if format.mode == "strict":
+            raise FormattingModeNotSupportedError(
+                formatting_mode="strict", provider="anthropic"
+            )
+        elif format.mode == "tool":
+            format_tool_schema = _formatting_utils.create_tool_schema(format)
+            anthropic_tools.append(_convert_tool_to_tool_param(format_tool_schema))
+            if tools:
+                kwargs["tool_choice"] = {"type": "any"}
+            else:
+                kwargs["tool_choice"] = {
+                    "type": "tool",
+                    "name": FORMAT_TOOL_NAME,
+                    "disable_parallel_tool_use": True,
+                }
+
+        if format.formatting_instructions:
+            messages = _base_utils.add_system_instructions(
+                messages, format.formatting_instructions
+            )
+
+    if anthropic_tools:
+        kwargs["tools"] = anthropic_tools
+
+    system_message_content, remaining_messages = _base_utils.extract_system_message(
+        messages
+    )
+
+    kwargs["messages"] = [
+        _encode_message(remaining_message, model_id, encode_thoughts)
+        for remaining_message in remaining_messages
+    ]
+
+    if system_message_content:
+        kwargs["system"] = system_message_content
+
+    return messages, format, kwargs