graphiti/graphiti_core/llm_client/anthropic_client.py

"""
Copyright 2024, Zep Software, Inc.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
"""

import json
import logging
import os
import typing
from json import JSONDecodeError
from typing import TYPE_CHECKING, Literal

from pydantic import BaseModel, ValidationError

from ..prompts.models import Message
from .client import LLMClient
from .config import DEFAULT_MAX_TOKENS, LLMConfig, ModelSize
from .errors import RateLimitError, RefusalError

if TYPE_CHECKING:
    import anthropic
    from anthropic import AsyncAnthropic
    from anthropic.types import MessageParam, ToolChoiceParam, ToolUnionParam
else:
    try:
        import anthropic
        from anthropic import AsyncAnthropic
        from anthropic.types import MessageParam, ToolChoiceParam, ToolUnionParam
    except ImportError:
        raise ImportError(
            'anthropic is required for AnthropicClient. '
            'Install it with: pip install graphiti-core[anthropic]'
        ) from None


logger = logging.getLogger(__name__)

AnthropicModel = Literal[
    'claude-sonnet-4-5-latest',
    'claude-sonnet-4-5-20250929',
    'claude-haiku-4-5-latest',
    'claude-3-7-sonnet-latest',
    'claude-3-7-sonnet-20250219',
    'claude-3-5-haiku-latest',
    'claude-3-5-haiku-20241022',
    'claude-3-5-sonnet-latest',
    'claude-3-5-sonnet-20241022',
    'claude-3-5-sonnet-20240620',
    'claude-3-opus-latest',
    'claude-3-opus-20240229',
    'claude-3-sonnet-20240229',
    'claude-3-haiku-20240307',
    'claude-2.1',
    'claude-2.0',
]

DEFAULT_MODEL: AnthropicModel = 'claude-haiku-4-5-latest'

# Maximum output tokens for different Anthropic models
# Based on official Anthropic documentation (as of 2025)
# Note: These represent standard limits without beta headers.
# Some models support higher limits with additional configuration (e.g., Claude 3.7 supports
# 128K with 'anthropic-beta: output-128k-2025-02-19' header, but this is not currently implemented).
ANTHROPIC_MODEL_MAX_TOKENS = {
    # Claude 4.5 models - 64K tokens
    'claude-sonnet-4-5-latest': 65536,
    'claude-sonnet-4-5-20250929': 65536,
    'claude-haiku-4-5-latest': 65536,
    # Claude 3.7 models - standard 64K tokens
    'claude-3-7-sonnet-latest': 65536,
    'claude-3-7-sonnet-20250219': 65536,
    # Claude 3.5 models
    'claude-3-5-haiku-latest': 8192,
    'claude-3-5-haiku-20241022': 8192,
    'claude-3-5-sonnet-latest': 8192,
    'claude-3-5-sonnet-20241022': 8192,
    'claude-3-5-sonnet-20240620': 8192,
    # Claude 3 models - 4K tokens
    'claude-3-opus-latest': 4096,
    'claude-3-opus-20240229': 4096,
    'claude-3-sonnet-20240229': 4096,
    'claude-3-haiku-20240307': 4096,
    # Claude 2 models - 4K tokens
    'claude-2.1': 4096,
    'claude-2.0': 4096,
}

# Default max tokens for models not in the mapping
DEFAULT_ANTHROPIC_MAX_TOKENS = 8192


class AnthropicClient(LLMClient):
    """
    A client for the Anthropic LLM.

    Args:
        config: A configuration object for the LLM.
        cache: Whether to cache the LLM responses.
        client: An optional client instance to use.
        max_tokens: The maximum number of tokens to generate.

    Methods:
        generate_response: Generate a response from the LLM.

    Notes:
        - If a LLMConfig is not provided, api_key will be pulled from the ANTHROPIC_API_KEY environment
            variable, and all default values will be used for the LLMConfig.

    """

    model: AnthropicModel

    def __init__(
        self,
        config: LLMConfig | None = None,
        cache: bool = False,
        client: AsyncAnthropic | None = None,
        max_tokens: int = DEFAULT_MAX_TOKENS,
    ) -> None:
        if config is None:
            config = LLMConfig()
            config.api_key = os.getenv('ANTHROPIC_API_KEY')
            config.max_tokens = max_tokens

        if config.model is None:
            config.model = DEFAULT_MODEL

        super().__init__(config, cache)
        # Explicitly set the instance model to the config model to prevent type checking errors
        self.model = typing.cast(AnthropicModel, config.model)

        if not client:
            self.client = AsyncAnthropic(
                api_key=config.api_key,
                max_retries=1,
            )
        else:
            self.client = client

    def _extract_json_from_text(self, text: str) -> dict[str, typing.Any]:
        """Extract JSON from text content.

        A helper method to extract JSON from text content, used when tool use fails or
        no response_model is provided.

        Args:
            text: The text to extract JSON from

        Returns:
            Extracted JSON as a dictionary

        Raises:
            ValueError: If JSON cannot be extracted or parsed
        """
        try:
            json_start = text.find('{')
            json_end = text.rfind('}') + 1
            if json_start >= 0 and json_end > json_start:
                json_str = text[json_start:json_end]
                return json.loads(json_str)
            else:
                raise ValueError(f'Could not extract JSON from model response: {text}')
        except (JSONDecodeError, ValueError) as e:
            raise ValueError(f'Could not extract JSON from model response: {text}') from e

    def _create_tool(
        self, response_model: type[BaseModel] | None = None
    ) -> tuple[list[ToolUnionParam], ToolChoiceParam]:
        """
        Create a tool definition based on the response_model if provided, or a generic JSON tool if not.

        Args:
            response_model: Optional Pydantic model to use for structured output.

        Returns:
            A list containing a single tool definition for use with the Anthropic API.
        """
        if response_model is not None:
            # Use the response_model to define the tool
            model_schema = response_model.model_json_schema()
            tool_name = response_model.__name__
            description = model_schema.get('description', f'Extract {tool_name} information')
        else:
            # Create a generic JSON output tool
            tool_name = 'generic_json_output'
            description = 'Output data in JSON format'
            model_schema = {
                'type': 'object',
                'additionalProperties': True,
                'description': 'Any JSON object containing the requested information',
            }

        tool = {
            'name': tool_name,
            'description': description,
            'input_schema': model_schema,
        }
        tool_list = [tool]
        tool_list_cast = typing.cast(list[ToolUnionParam], tool_list)
        tool_choice = {'type': 'tool', 'name': tool_name}
        tool_choice_cast = typing.cast(ToolChoiceParam, tool_choice)
        return tool_list_cast, tool_choice_cast

    def _get_max_tokens_for_model(self, model: str) -> int:
        """Get the maximum output tokens for a specific Anthropic model.

        Args:
            model: The model name to look up

        Returns:
            int: The maximum output tokens for the model
        """
        return ANTHROPIC_MODEL_MAX_TOKENS.get(model, DEFAULT_ANTHROPIC_MAX_TOKENS)

    def _resolve_max_tokens(self, requested_max_tokens: int | None, model: str) -> int:
        """
        Resolve the maximum output tokens to use based on precedence rules.

        Precedence order (highest to lowest):
        1. Explicit max_tokens parameter passed to generate_response()
        2. Instance max_tokens set during client initialization
        3. Model-specific maximum tokens from ANTHROPIC_MODEL_MAX_TOKENS mapping
        4. DEFAULT_ANTHROPIC_MAX_TOKENS as final fallback

        Args:
            requested_max_tokens: The max_tokens parameter passed to generate_response()
            model: The model name to look up model-specific limits

        Returns:
            int: The resolved maximum tokens to use
        """
        # 1. Use explicit parameter if provided
        if requested_max_tokens is not None:
            return requested_max_tokens

        # 2. Use instance max_tokens if set during initialization
        if self.max_tokens is not None:
            return self.max_tokens

        # 3. Use model-specific maximum or return DEFAULT_ANTHROPIC_MAX_TOKENS
        return self._get_max_tokens_for_model(model)

    async def _generate_response(
        self,
        messages: list[Message],
        response_model: type[BaseModel] | None = None,
        max_tokens: int | None = None,
        model_size: ModelSize = ModelSize.medium,
    ) -> dict[str, typing.Any]:
        """
        Generate a response from the Anthropic LLM using tool-based approach for all requests.

        Args:
            messages: List of message objects to send to the LLM.
            response_model: Optional Pydantic model to use for structured output.
            max_tokens: Maximum number of tokens to generate.

        Returns:
            Dictionary containing the structured response from the LLM.

        Raises:
            RateLimitError: If the rate limit is exceeded.
            RefusalError: If the LLM refuses to respond.
            Exception: If an error occurs during the generation process.
        """
        system_message = messages[0]
        user_messages = [{'role': m.role, 'content': m.content} for m in messages[1:]]
        user_messages_cast = typing.cast(list[MessageParam], user_messages)

        # Resolve max_tokens dynamically based on the model's capabilities
        # This allows different models to use their full output capacity
        max_creation_tokens: int = self._resolve_max_tokens(max_tokens, self.model)

        try:
            # Create the appropriate tool based on whether response_model is provided
            tools, tool_choice = self._create_tool(response_model)
            result = await self.client.messages.create(
                system=system_message.content,
                max_tokens=max_creation_tokens,
                temperature=self.temperature,
                messages=user_messages_cast,
                model=self.model,
                tools=tools,
                tool_choice=tool_choice,
            )

            # Extract the tool output from the response
            for content_item in result.content:
                if content_item.type == 'tool_use':
                    if isinstance(content_item.input, dict):
                        tool_args: dict[str, typing.Any] = content_item.input
                    else:
                        tool_args = json.loads(str(content_item.input))
                    return tool_args

            # If we didn't get a proper tool_use response, try to extract from text
            for content_item in result.content:
                if content_item.type == 'text':
                    return self._extract_json_from_text(content_item.text)
                else:
                    raise ValueError(
                        f'Could not extract structured data from model response: {result.content}'
                    )

            # If we get here, we couldn't parse a structured response
            raise ValueError(
                f'Could not extract structured data from model response: {result.content}'
            )

        except anthropic.RateLimitError as e:
            raise RateLimitError(f'Rate limit exceeded. Please try again later. Error: {e}') from e
        except anthropic.APIError as e:
            # Special case for content policy violations. We convert these to RefusalError
            # to bypass the retry mechanism, as retrying policy-violating content will always fail.
            # This avoids wasting API calls and provides more specific error messaging to the user.
            if 'refused to respond' in str(e).lower():
                raise RefusalError(str(e)) from e
            raise e
        except Exception as e:
            raise e

    async def generate_response(
        self,
        messages: list[Message],
        response_model: type[BaseModel] | None = None,
        max_tokens: int | None = None,
        model_size: ModelSize = ModelSize.medium,
        group_id: str | None = None,
        prompt_name: str | None = None,
    ) -> dict[str, typing.Any]:
        """
        Generate a response from the LLM.

        Args:
            messages: List of message objects to send to the LLM.
            response_model: Optional Pydantic model to use for structured output.
            max_tokens: Maximum number of tokens to generate.

        Returns:
            Dictionary containing the structured response from the LLM.

        Raises:
            RateLimitError: If the rate limit is exceeded.
            RefusalError: If the LLM refuses to respond.
            Exception: If an error occurs during the generation process.
        """
        if max_tokens is None:
            max_tokens = self.max_tokens

        # Wrap entire operation in tracing span
        with self.tracer.start_span('llm.generate') as span:
            attributes = {
                'llm.provider': 'anthropic',
                'model.size': model_size.value,
                'max_tokens': max_tokens,
            }
            if prompt_name:
                attributes['prompt.name'] = prompt_name
            span.add_attributes(attributes)

            retry_count = 0
            max_retries = 2
            last_error: Exception | None = None

            while retry_count <= max_retries:
                try:
                    response = await self._generate_response(
                        messages, response_model, max_tokens, model_size
                    )

                    # If we have a response_model, attempt to validate the response
                    if response_model is not None:
                        # Validate the response against the response_model
                        model_instance = response_model(**response)
                        return model_instance.model_dump()

                    # If no validation needed, return the response
                    return response

                except (RateLimitError, RefusalError):
                    # These errors should not trigger retries
                    span.set_status('error', str(last_error))
                    raise
                except Exception as e:
                    last_error = e

                    if retry_count >= max_retries:
                        if isinstance(e, ValidationError):
                            logger.error(
                                f'Validation error after {retry_count}/{max_retries} attempts: {e}'
                            )
                        else:
                            logger.error(f'Max retries ({max_retries}) exceeded. Last error: {e}')
                        span.set_status('error', str(e))
                        span.record_exception(e)
                        raise e

                    if isinstance(e, ValidationError):
                        response_model_cast = typing.cast(type[BaseModel], response_model)
                        error_context = f'The previous response was invalid. Please provide a valid {response_model_cast.__name__} object. Error: {e}'
                    else:
                        error_context = (
                            f'The previous response attempt was invalid. '
                            f'Error type: {e.__class__.__name__}. '
                            f'Error details: {str(e)}. '
                            f'Please try again with a valid response.'
                        )

                    # Common retry logic
                    retry_count += 1
                    messages.append(Message(role='user', content=error_context))
                    logger.warning(
                        f'Retrying after error (attempt {retry_count}/{max_retries}): {e}'
                    )

            # If we somehow get here, raise the last error
            span.set_status('error', str(last_error))
            raise last_error or Exception('Max retries exceeded with no specific error')