mirascope 1.25.6__py3-none-any.whl → 2.0.0__py3-none-any.whl

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

@@ -0,0 +1,288 @@
+"""Standard 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 anthropic.types.beta import BetaUsage
+
+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,
+    Usage,
+    UsageDeltaChunk,
+)
+from ..model_id import AnthropicModelId, model_name
+
+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_usage(
+    usage: anthropic_types.Usage | BetaUsage,
+) -> Usage:
+    """Convert Anthropic Usage (or BetaUsage) to Mirascope Usage."""
+
+    cache_read_tokens = usage.cache_read_input_tokens or 0
+    cache_write_tokens = usage.cache_creation_input_tokens or 0
+    input_tokens = usage.input_tokens + cache_read_tokens + cache_write_tokens
+    output_tokens = usage.output_tokens
+    return Usage(
+        input_tokens=input_tokens,
+        output_tokens=output_tokens,
+        cache_read_tokens=cache_read_tokens,
+        cache_write_tokens=cache_write_tokens,
+        reasoning_tokens=0,
+        raw=usage,
+    )
+
+
+def decode_response(
+    response: anthropic_types.Message,
+    model_id: AnthropicModelId,
+    *,
+    include_thoughts: bool,
+) -> tuple[AssistantMessage, FinishReason | None, Usage]:
+    """Convert Anthropic message to mirascope AssistantMessage and usage."""
+    content = [_decode_assistant_content(part) for part in response.content]
+    if not include_thoughts:
+        content = [part for part in content if part.type != "thought"]
+    assistant_message = AssistantMessage(
+        content=content,
+        provider_id="anthropic",
+        model_id=model_id,
+        provider_model_name=model_name(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
+    )
+    usage = decode_usage(response.usage)
+    return assistant_message, finish_reason, usage
+
+
+ContentBlock: TypeAlias = (
+    anthropic_types.TextBlockParam
+    | anthropic_types.ThinkingBlockParam
+    | anthropic_types.ToolUseBlockParam
+    | anthropic_types.RedactedThinkingBlockParam
+)
+
+
+class _AnthropicChunkProcessor:
+    """Processes Anthropic stream events and maintains state across events."""
+
+    def __init__(self, *, include_thoughts: bool) -> None:
+        self.current_block_param: ContentBlock | None = None
+        self.accumulated_tool_json: str = ""
+        self.accumulated_blocks: list[ContentBlock] = []
+        self.include_thoughts = include_thoughts
+
+    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": "",
+                }
+                if self.include_thoughts:
+                    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(
+                    id=self.current_block_param["id"], 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
+                if self.include_thoughts:
+                    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(id=self.current_block_param["id"])
+            elif block_type == "thinking":
+                if self.include_thoughts:
+                    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)
+
+            # Emit usage delta
+            usage = event.usage
+            yield UsageDeltaChunk(
+                input_tokens=usage.input_tokens or 0,
+                output_tokens=usage.output_tokens,
+                cache_read_tokens=usage.cache_read_input_tokens or 0,
+                cache_write_tokens=usage.cache_creation_input_tokens or 0,
+                reasoning_tokens=0,
+            )
+
+    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, *, include_thoughts: bool
+) -> ChunkIterator:
+    """Returns a ChunkIterator converted from an Anthropic MessageStreamManager."""
+    processor = _AnthropicChunkProcessor(include_thoughts=include_thoughts)
+    with anthropic_stream_manager as stream:
+        for event in stream._raw_stream:  # pyright: ignore[reportPrivateUsage]
+            yield from processor.process_event(event)
+    yield processor.raw_message_chunk()
+
+
+async def decode_async_stream(
+    anthropic_stream_manager: AsyncMessageStreamManager, *, include_thoughts: bool
+) -> AsyncChunkIterator:
+    """Returns an AsyncChunkIterator converted from an Anthropic MessageStreamManager."""
+    processor = _AnthropicChunkProcessor(include_thoughts=include_thoughts)
+    async with anthropic_stream_manager as stream:
+        async for event in stream._raw_stream:  # pyright: ignore[reportPrivateUsage]
+            for item in processor.process_event(event):
+                yield item
+    yield processor.raw_message_chunk()

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

@@ -0,0 +1,418 @@
+"""Shared Anthropic encoding utilities."""
+
+from __future__ import annotations
+
+import json
+from collections.abc import Sequence
+from functools import lru_cache
+from typing import TYPE_CHECKING, Any, 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
+from ....formatting import (
+    Format,
+    FormattableT,
+    OutputParser,
+    resolve_format,
+)
+from ....messages import AssistantMessage, Message, UserMessage
+from ....tools import FORMAT_TOOL_NAME, AnyToolSchema, BaseToolkit
+from ...base import _utils as _base_utils
+from ..model_id import AnthropicModelId, model_name
+
+if TYPE_CHECKING:
+    from ....models import Params, ThinkingLevel
+
+DEFAULT_MAX_TOKENS = 16000
+# TODO: Change DEFAULT_FORMAT_MODE to strict when strict is no longer a beta feature.
+DEFAULT_FORMAT_MODE = "tool"
+
+# Thinking level to a float multiplier % of max tokens
+THINKING_LEVEL_TO_BUDGET_MULTIPLIER: dict[ThinkingLevel, float] = {
+    "minimal": 0,  # Will become 1024 (actual minimal value)
+    "low": 0.2,
+    "medium": 0.4,
+    "high": 0.6,
+    "max": 0.8,
+}
+
+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_id="anthropic"
+    )  # pragma: no cover
+
+
+def compute_thinking_budget(
+    level: ThinkingLevel,
+    max_tokens: int,
+) -> int:
+    """Compute Anthropic token budget from ThinkingConfig level.
+
+    Args:
+        level: The thinking level from ThinkingConfig
+        max_tokens: The max_tokens value for the request
+
+    Returns:
+        Token budget for thinking (0 to disable, positive for budget)
+    """
+
+    if level == "none":
+        return 0
+    elif level == "default":
+        return -1  # Do not set thinking, leave to provider default
+
+    multiplier: float = THINKING_LEVEL_TO_BUDGET_MULTIPLIER.get(level, 0.4)
+    budget = int(multiplier * max_tokens)
+    return max(1024, budget)  # Always return at least 1024, minimum allowed budget
+
+
+class ProcessedParams(TypedDict, total=False):
+    """Common parameters processed from Params."""
+
+    temperature: float
+    max_tokens: int
+    top_p: float
+    top_k: int
+    stop_sequences: list[str]
+    thinking: dict[str, Any]
+    encode_thoughts_as_text: bool
+
+
+def process_params(params: Params, default_max_tokens: int) -> ProcessedParams:
+    """Process common Anthropic parameters from Params.
+
+    Returns a dict with processed parameters that can be merged into kwargs.
+    """
+    result: ProcessedParams = {
+        "max_tokens": default_max_tokens,
+        "encode_thoughts_as_text": False,
+    }
+
+    with _base_utils.ensure_all_params_accessed(
+        params=params, provider_id="anthropic", unsupported_params=["seed"]
+    ) as param_accessor:
+        if param_accessor.temperature is not None:
+            result["temperature"] = param_accessor.temperature
+        if param_accessor.max_tokens is not None:
+            result["max_tokens"] = param_accessor.max_tokens
+        if param_accessor.top_p is not None:
+            result["top_p"] = param_accessor.top_p
+        if param_accessor.top_k is not None:
+            result["top_k"] = param_accessor.top_k
+        if param_accessor.stop_sequences is not None:
+            result["stop_sequences"] = param_accessor.stop_sequences
+        if param_accessor.thinking is not None:
+            thinking_config = param_accessor.thinking
+            level = thinking_config.get("level")
+
+            # Compute token budget from level
+            budget_tokens = compute_thinking_budget(level, result["max_tokens"])
+            if budget_tokens == 0:
+                result["thinking"] = {"type": "disabled"}
+            elif budget_tokens > 0:
+                result["thinking"] = {"type": "enabled", "budget_tokens": budget_tokens}
+            else:
+                # budget is -1, do not set thinking at all.
+                pass
+
+            # Handle encode_thoughts_as_text from ThinkingConfig
+            if thinking_config.get("encode_thoughts_as_text"):
+                result["encode_thoughts_as_text"] = True
+
+    return result
+
+
+class MessageCreateKwargs(TypedDict, total=False):
+    """Kwargs for Anthropic Message.create method."""
+
+    model: Required[str]
+    max_tokens: Required[int]
+    messages: Sequence[anthropic_types.MessageParam]
+    system: Sequence[anthropic_types.TextBlockParam] | 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,
+    add_cache_control: bool,
+) -> str | Sequence[anthropic_types.ContentBlockParam]:
+    """Convert mirascope content to Anthropic content format."""
+
+    if len(content) == 1 and content[0].type == "text":
+        if not content[0].text:
+            raise FeatureNotSupportedError(
+                "empty message content",
+                "anthropic",
+                message="Anthropic does not support empty message content.",
+            )
+        if add_cache_control:
+            return [
+                anthropic_types.TextBlockParam(
+                    type="text",
+                    text=content[0].text,
+                    cache_control={"type": "ephemeral"},
+                )
+            ]
+        return content[0].text
+
+    blocks: list[anthropic_types.ContentBlockParam] = []
+
+    # Find the last cacheable content part (text, image, tool_result, or tool_call)
+    last_cacheable_index = -1
+    if add_cache_control:
+        for i in range(len(content) - 1, -1, -1):
+            part = content[i]
+            if part.type in ("text", "image", "tool_output", "tool_call"):
+                if part.type == "text" and not part.text:  # pragma: no cover
+                    continue  # Skip empty text
+                last_cacheable_index = i
+                break
+
+    for i, part in enumerate(content):
+        should_add_cache = add_cache_control and i == last_cacheable_index
+
+        if part.type == "text":
+            if part.text:
+                blocks.append(
+                    anthropic_types.TextBlockParam(
+                        type="text",
+                        text=part.text,
+                        cache_control={"type": "ephemeral"}
+                        if should_add_cache
+                        else None,
+                    )
+                )
+        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,
+                    cache_control={"type": "ephemeral"} if should_add_cache else None,
+                )
+            )
+        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.result),
+                    cache_control={"type": "ephemeral"} if should_add_cache else None,
+                )
+            )
+        elif part.type == "tool_call":
+            blocks.append(
+                anthropic_types.ToolUseBlockParam(
+                    type="tool_use",
+                    id=part.id,
+                    name=part.name,
+                    input=json.loads(part.args),
+                    cache_control={"type": "ephemeral"} if should_add_cache else None,
+                )
+            )
+        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}")
+
+    if not blocks:
+        raise FeatureNotSupportedError(
+            "empty message content",
+            "anthropic",
+            message="Anthropic does not support empty message content.",
+        )
+
+    return blocks
+
+
+def _encode_message(
+    message: UserMessage | AssistantMessage,
+    model_id: AnthropicModelId,
+    encode_thoughts: bool,
+    add_cache_control: bool = False,
+) -> anthropic_types.MessageParam:
+    """Convert user or assistant Message to Anthropic MessageParam format.
+
+    Args:
+        message: The message to encode
+        model_id: The Anthropic model ID
+        encode_thoughts: Whether to encode thought blocks as text
+        add_cache_control: Whether to add cache_control to the last content block
+    """
+    if (
+        message.role == "assistant"
+        and message.provider_id == "anthropic"
+        and message.model_id == model_id
+        and message.raw_message
+        and not encode_thoughts
+        and not add_cache_control
+    ):
+        return cast(anthropic_types.MessageParam, message.raw_message)
+
+    content = encode_content(message.content, encode_thoughts, add_cache_control)
+
+    return {
+        "role": message.role,
+        "content": content,
+    }
+
+
+def _encode_messages(
+    messages: Sequence[UserMessage | AssistantMessage],
+    model_id: AnthropicModelId,
+    encode_thoughts_as_text: bool,
+) -> Sequence[anthropic_types.MessageParam]:
+    """Encode messages and add cache control for multi-turn conversations.
+
+    If the conversation contains assistant messages (indicating multi-turn),
+    adds cache_control to the last content block of the last message.
+    """
+    # Detect multi-turn conversations by checking for assistant messages
+    has_assistant_message = any(msg.role == "assistant" for msg in messages)
+
+    # Encode messages, adding cache_control to the last message if multi-turn
+    encoded_messages: list[anthropic_types.MessageParam] = []
+    for i, message in enumerate(messages):
+        is_last = i == len(messages) - 1
+        add_cache = has_assistant_message and is_last
+        encoded_messages.append(
+            _encode_message(message, model_id, encode_thoughts_as_text, add_cache)
+        )
+    return encoded_messages
+
+
+@lru_cache(maxsize=128)
+def convert_tool_to_tool_param(tool: AnyToolSchema) -> 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[AnyToolSchema] | BaseToolkit[AnyToolSchema] | None,
+    format: type[FormattableT]
+    | Format[FormattableT]
+    | OutputParser[FormattableT]
+    | None,
+    params: Params,
+) -> tuple[Sequence[Message], Format[FormattableT] | None, MessageCreateKwargs]:
+    """Prepares a request for the Anthropic messages.create method."""
+
+    processed = process_params(params, DEFAULT_MAX_TOKENS)
+    encode_thoughts_as_text = processed.pop("encode_thoughts_as_text", False)
+    max_tokens = processed.pop("max_tokens", DEFAULT_MAX_TOKENS)
+
+    kwargs: MessageCreateKwargs = MessageCreateKwargs(
+        {"model": model_name(model_id), "max_tokens": max_tokens, **processed}
+    )
+
+    tools = tools.tools if isinstance(tools, BaseToolkit) else tools or []
+
+    # Check for strict tools - the non-beta API doesn't support them
+    if _base_utils.has_strict_tools(tools):
+        raise FeatureNotSupportedError(
+            feature="strict tools",
+            provider_id="anthropic",
+            model_id=model_id,
+            message="Anthropic provider does not support strict tools. Try the beta provider.",
+        )
+
+    anthropic_tools = [convert_tool_to_tool_param(tool) for tool in tools]
+    format = resolve_format(format, default_mode=DEFAULT_FORMAT_MODE)
+    if format is not None:
+        if format.mode == "strict":
+            raise FeatureNotSupportedError(
+                feature="formatting_mode:strict",
+                provider_id="anthropic",
+                model_id=model_id,
+            )
+        if format.mode == "tool":
+            format_tool_schema = format.create_tool_schema()
+            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:
+        # Add cache control to the last tool for prompt caching
+        last_tool = anthropic_tools[-1]
+        last_tool["cache_control"] = {"type": "ephemeral"}
+        kwargs["tools"] = anthropic_tools
+
+    system_message_content, remaining_messages = _base_utils.extract_system_message(
+        messages
+    )
+
+    kwargs["messages"] = _encode_messages(
+        remaining_messages, model_id, encode_thoughts_as_text
+    )
+
+    if system_message_content:
+        kwargs["system"] = [
+            anthropic_types.TextBlockParam(
+                type="text",
+                text=system_message_content,
+                cache_control={"type": "ephemeral"},
+            )
+        ]
+
+    return messages, format, kwargs