Implementation:Arize ai Phoenix Legacy AnthropicModel

Overview

AnthropicModel is a legacy model wrapper in the phoenix-evals package that provides an interface for using Anthropic Claude models within Phoenix LLM evaluations. It extends BaseModel and integrates with the Anthropic SDK, providing both synchronous and asynchronous generation, dynamic rate limiting that adjusts based on API throttling responses, and structured token usage tracking. The wrapper formats prompts into the Anthropic Messages API format and handles context limit errors gracefully.

LLM_Evaluation Model_Integration

Description

The AnthropicModel class is implemented as a Python dataclass that extends the abstract BaseModel. Key characteristics include:

• Anthropic SDK integration: Initializes both synchronous (Anthropic()) and asynchronous (AsyncAnthropic()) clients from the anthropic package.
• Version validation: Enforces a minimum Anthropic SDK version of 0.18.0 at initialization time.
• Dynamic rate limiting: Uses the RateLimiter infrastructure, configured to intercept anthropic.RateLimitError with an initial rate of 5 requests per second and a 1-minute enforcement window.
• Context limit handling: Catches BadRequestError exceptions containing "prompt is too long" messages and re-raises them as PhoenixContextLimitExceeded.
• Full async support: Both _generate_with_extra() and _async_generate_with_extra() are natively implemented.
• Tool use support: The _extract_text() method detects tool use blocks in responses and serializes them as JSON.
• Cache-aware usage tracking: Token usage extraction accounts for cache_creation_input_tokens and cache_read_input_tokens in addition to standard input_tokens.

Usage

# Set the ANTHROPIC_API_KEY environment variable before use

from phoenix.evals.models import AnthropicModel

# Basic usage with defaults (claude-2.1)
model = AnthropicModel()

# Specify a different model and parameters
model = AnthropicModel(
    model="claude-3-5-sonnet-20241022",
    temperature=0.5,
    max_tokens=2048,
    top_p=0.9,
    initial_rate_limit=10,
)

# Direct invocation
response = model("Explain quantum computing in simple terms.")
print(response)

Code Reference

Source Location

Property	Value
Repository	Arize-ai/phoenix
File	packages/phoenix-evals/src/phoenix/evals/legacy/models/anthropic.py
Lines	226
Module	phoenix.evals.legacy.models.anthropic

Class Signature

@dataclass
class AnthropicModel(BaseModel):
    model: str = "claude-2.1"
    temperature: float = 0.0
    max_tokens: int = 1024
    top_p: float = 1
    top_k: int = 256
    stop_sequences: List[str] = field(default_factory=list)
    extra_parameters: Dict[str, Any] = field(default_factory=dict)
    max_content_size: Optional[int] = None
    initial_rate_limit: int = 5
    timeout: int = 120

Constructor Parameters

Parameter	Type	Default	Description
model	str	"claude-2.1"	The Anthropic model name to use.
temperature	float	0.0	Sampling temperature for generation.
max_tokens	int	1024	Maximum number of tokens to generate.
top_p	float	1	Nucleus sampling probability mass.
top_k	int	256	Top-K sampling cutoff.
stop_sequences	List[str]	[]	Sequences that halt generation.
extra_parameters	Dict[str, Any]	{}	Extra parameters for the request body (e.g., countPenalty for A21 models).
max_content_size	Optional[int]	None	Maximum content size for fine-tuned models.
initial_rate_limit	int	5	Initial requests-per-second rate limit.
timeout	int	120	Timeout for API requests in seconds.

Key Methods

Method	Signature	Description
__post_init__	(self) -> None	Initializes the Anthropic client and rate limiter.
_init_client	(self) -> None	Validates SDK version and creates sync/async Anthropic clients.
_init_rate_limiter	(self) -> None	Configures the rate limiter with RateLimitError.
invocation_parameters	(self) -> Dict[str, Any]	Returns the API call parameters dictionary.
_generate_with_extra	(self, prompt, **kwargs) -> Tuple[str, ExtraInfo]	Synchronous generation with rate limiting.
_async_generate_with_extra	async (self, prompt, **kwargs) -> Tuple[str, ExtraInfo]	Asynchronous generation with rate limiting.
_format_prompt_for_claude	(self, prompt: MultimodalPrompt) -> List[Dict[str, str]]	Converts a MultimodalPrompt to Anthropic Messages API format.
_extract_text	(self, message: Message) -> str	Extracts text or tool-use JSON from an Anthropic response.
_extract_usage	(self, message_usage: MessageUsage) -> Usage	Extracts token usage including cache tokens.
_parse_output	(self, message: Message) -> Tuple[str, ExtraInfo]	Combines text extraction and usage extraction.

Import

from phoenix.evals.models import AnthropicModel

I/O Contract

Direction	Type	Description
Input	Union[str, MultimodalPrompt]	A text string or multimodal prompt (only text parts supported for Anthropic).
Input (optional)	Optional[str]	Instruction parameter (ignored; stripped before API call).
Output	str	Generated text response, or JSON-serialized tool use arguments.
Output (with extra)	Tuple[str, ExtraInfo]	Generated text paired with ExtraInfo containing Usage token counts.
Error	PhoenixContextLimitExceeded	Raised when prompt exceeds the model's context window.
Error	ImportError	Raised if anthropic package is not installed or version is too old.

Usage Examples

Basic Generation

from phoenix.evals.models import AnthropicModel

model = AnthropicModel(model="claude-3-5-sonnet-20241022", temperature=0.0)
response = model("What are the benefits of test-driven development?")
print(response)

Async Generation

import asyncio
from phoenix.evals.models import AnthropicModel

model = AnthropicModel(model="claude-3-5-sonnet-20241022")

async def generate():
    result = await model._async_generate("Summarize the theory of relativity.")
    return result

response = asyncio.run(generate())
print(response)

With Custom Rate Limiting

from phoenix.evals.models import AnthropicModel

# Higher rate limit for batch processing
model = AnthropicModel(
    model="claude-3-haiku-20240307",
    initial_rate_limit=20,
    max_tokens=512,
)

Related Pages

• Implementation:Arize_ai_Phoenix_Legacy_BaseModel - Abstract base class that AnthropicModel extends
• Implementation:Arize_ai_Phoenix_Legacy_OpenAIModel - OpenAI model wrapper (similar pattern)
• Implementation:Arize_ai_Phoenix_Legacy_MistralAIModel - Mistral AI model wrapper (similar pattern)
• Implementation:Arize_ai_Phoenix_Legacy_BedrockModel - AWS Bedrock wrapper (supports Anthropic models via Bedrock)
• Implementation:Arize_ai_Phoenix_Legacy_GoogleGenAIModel - Google GenAI model wrapper