judgeval 0.16.0__py3-none-any.whl → 0.16.1__py3-none-any.whl

judgeval/tracer/llm/__init__.py

@@ -1,1264 +1,7 @@
 from __future__ import annotations
-import functools
-from typing import (
-    Tuple,
-    Optional,
-    Any,
-    TYPE_CHECKING,
-    Union,
-    AsyncGenerator,
-    Generator,
-    Iterator,
-    AsyncIterator,
-)
-from functools import wraps
-from enum import Enum
-from judgeval.data.trace import TraceUsage
-from judgeval.logger import judgeval_logger
-from litellm.cost_calculator import cost_per_token as _original_cost_per_token
-from opentelemetry.trace import Span
 
-from judgeval.tracer.llm.providers import (
-    HAS_OPENAI,
-    HAS_TOGETHER,
-    HAS_ANTHROPIC,
-    HAS_GOOGLE_GENAI,
-    HAS_GROQ,
-    ApiClient,
-)
-from judgeval.tracer.managers import sync_span_context, async_span_context
-from judgeval.tracer.keys import AttributeKeys
-from judgeval.utils.serialize import safe_serialize
-from judgeval.tracer.utils import set_span_attribute
 
-if TYPE_CHECKING:
-    from judgeval.tracer import Tracer
+from .config import _detect_provider, wrap_provider
 
 
-class ProviderType(Enum):
-    """Enum for different LLM provider types."""
-
-    OPENAI = "openai"
-    ANTHROPIC = "anthropic"
-    TOGETHER = "together"
-    GOOGLE = "google"
-    GROQ = "groq"
-    DEFAULT = "default"
-
-
-@wraps(_original_cost_per_token)
-def cost_per_token(
-    *args: Any, **kwargs: Any
-) -> Tuple[Optional[float], Optional[float]]:
-    try:
-        prompt_tokens_cost_usd_dollar, completion_tokens_cost_usd_dollar = (
-            _original_cost_per_token(*args, **kwargs)
-        )
-        if (
-            prompt_tokens_cost_usd_dollar == 0
-            and completion_tokens_cost_usd_dollar == 0
-        ):
-            judgeval_logger.warning("LiteLLM returned a total of 0 for cost per token")
-        return prompt_tokens_cost_usd_dollar, completion_tokens_cost_usd_dollar
-    except Exception as e:
-        judgeval_logger.warning(f"Error calculating cost per token: {e}")
-        return None, None
-
-
-def _detect_provider(client: ApiClient) -> ProviderType:
-    """Detect the provider type of the client once to avoid repeated isinstance checks."""
-    if HAS_OPENAI:
-        from judgeval.tracer.llm.providers import openai_OpenAI, openai_AsyncOpenAI
-
-        assert openai_OpenAI is not None, "OpenAI client not found"
-        assert openai_AsyncOpenAI is not None, "OpenAI async client not found"
-        if isinstance(client, (openai_OpenAI, openai_AsyncOpenAI)):
-            return ProviderType.OPENAI
-
-    if HAS_ANTHROPIC:
-        from judgeval.tracer.llm.providers import (
-            anthropic_Anthropic,
-            anthropic_AsyncAnthropic,
-        )
-
-        assert anthropic_Anthropic is not None, "Anthropic client not found"
-        assert anthropic_AsyncAnthropic is not None, "Anthropic async client not found"
-        if isinstance(client, (anthropic_Anthropic, anthropic_AsyncAnthropic)):
-            return ProviderType.ANTHROPIC
-
-    if HAS_TOGETHER:
-        from judgeval.tracer.llm.providers import (
-            together_Together,
-            together_AsyncTogether,
-        )
-
-        assert together_Together is not None, "Together client not found"
-        assert together_AsyncTogether is not None, "Together async client not found"
-        if isinstance(client, (together_Together, together_AsyncTogether)):
-            return ProviderType.TOGETHER
-
-    if HAS_GOOGLE_GENAI:
-        from judgeval.tracer.llm.providers import (
-            google_genai_Client,
-            google_genai_AsyncClient,
-        )
-
-        assert google_genai_Client is not None, "Google GenAI client not found"
-        assert google_genai_AsyncClient is not None, (
-            "Google GenAI async client not found"
-        )
-        if isinstance(client, (google_genai_Client, google_genai_AsyncClient)):
-            return ProviderType.GOOGLE
-
-    if HAS_GROQ:
-        from judgeval.tracer.llm.providers import groq_Groq, groq_AsyncGroq
-
-        assert groq_Groq is not None, "Groq client not found"
-        assert groq_AsyncGroq is not None, "Groq async client not found"
-        if isinstance(client, (groq_Groq, groq_AsyncGroq)):
-            return ProviderType.GROQ
-
-    return ProviderType.DEFAULT
-
-
-# Provider-specific content extraction handlers
-def _extract_openai_content(chunk) -> str:
-    """Extract content from OpenAI streaming chunk."""
-    if (
-        hasattr(chunk, "choices")
-        and chunk.choices
-        and hasattr(chunk.choices[0], "delta")
-    ):
-        delta_content = getattr(chunk.choices[0].delta, "content", None)
-        if delta_content:
-            return delta_content
-    return ""
-
-
-def _extract_anthropic_content(chunk) -> str:
-    """Extract content from Anthropic streaming chunk."""
-    if hasattr(chunk, "type"):
-        if chunk.type == "content_block_delta":
-            if hasattr(chunk, "delta"):
-                if hasattr(chunk.delta, "text"):
-                    return chunk.delta.text or ""
-                elif hasattr(chunk.delta, "partial_json"):
-                    # Tool use input streaming - return raw JSON to accumulate properly
-                    return chunk.delta.partial_json or ""
-        elif chunk.type == "content_block_start":
-            if hasattr(chunk, "content_block") and hasattr(chunk.content_block, "type"):
-                if chunk.content_block.type == "tool_use":
-                    tool_info = {
-                        "type": "tool_use",
-                        "id": getattr(chunk.content_block, "id", None),
-                        "name": getattr(chunk.content_block, "name", None),
-                    }
-                    return f"[TOOL_USE_START: {tool_info}]"
-    elif hasattr(chunk, "delta") and hasattr(chunk.delta, "text"):
-        return chunk.delta.text or ""
-    elif hasattr(chunk, "text"):
-        return chunk.text or ""
-    return ""
-
-
-def _extract_together_content(chunk) -> str:
-    """Extract content from Together streaming chunk."""
-    if hasattr(chunk, "choices") and chunk.choices:
-        choice = chunk.choices[0]
-        if hasattr(choice, "delta") and hasattr(choice.delta, "content"):
-            return choice.delta.content or ""
-    return ""
-
-
-def _extract_groq_content(chunk) -> str:
-    """Extract content from Groq streaming chunk."""
-    if hasattr(chunk, "choices") and chunk.choices:
-        choice = chunk.choices[0]
-        if hasattr(choice, "delta") and hasattr(choice.delta, "content"):
-            return choice.delta.content or ""
-    return ""
-
-
-# Provider-specific chunk usage extraction handlers
-def _extract_openai_chunk_usage(chunk) -> Any:
-    """Extract usage data from OpenAI streaming chunk."""
-    if hasattr(chunk, "usage") and chunk.usage:
-        return chunk.usage
-    return None
-
-
-def _extract_anthropic_chunk_usage(chunk) -> Any:
-    """Extract usage data from Anthropic streaming chunk."""
-    if hasattr(chunk, "type"):
-        if chunk.type == "message_start":
-            if hasattr(chunk, "message") and hasattr(chunk.message, "usage"):
-                return chunk.message.usage
-        elif chunk.type == "message_delta":
-            if hasattr(chunk, "usage"):
-                return chunk.usage
-        elif chunk.type == "message_stop":
-            if hasattr(chunk, "usage"):
-                return chunk.usage
-    return None
-
-
-def _extract_together_chunk_usage(chunk) -> Any:
-    """Extract usage data from Together streaming chunk."""
-    if hasattr(chunk, "usage") and chunk.usage:
-        return chunk.usage
-    return None
-
-
-def _extract_groq_chunk_usage(chunk) -> Any:
-    """Extract usage data from Groq streaming chunk."""
-    # Groq provides usage data in the last chunk when stream_options={"include_usage": True} is used
-    if hasattr(chunk, "usage") and chunk.usage:
-        return chunk.usage
-    return None
-
-
-# Provider-specific token extraction handlers
-def _extract_openai_tokens(usage_data) -> tuple[int, int, int, int]:
-    """Extract token counts from OpenAI usage data."""
-    prompt_tokens = (
-        usage_data.prompt_tokens
-        if hasattr(usage_data, "prompt_tokens") and usage_data.prompt_tokens is not None
-        else 0
-    )
-    completion_tokens = (
-        usage_data.completion_tokens
-        if hasattr(usage_data, "completion_tokens")
-        and usage_data.completion_tokens is not None
-        else 0
-    )
-    return prompt_tokens, completion_tokens, 0, 0
-
-
-def _extract_anthropic_tokens(usage_data) -> tuple[int, int, int, int]:
-    """Extract token counts from Anthropic usage data."""
-    prompt_tokens = (
-        usage_data.input_tokens
-        if hasattr(usage_data, "input_tokens") and usage_data.input_tokens is not None
-        else 0
-    )
-    completion_tokens = (
-        usage_data.output_tokens
-        if hasattr(usage_data, "output_tokens") and usage_data.output_tokens is not None
-        else 0
-    )
-    cache_read_input_tokens = (
-        usage_data.cache_read_input_tokens
-        if hasattr(usage_data, "cache_read_input_tokens")
-        and usage_data.cache_read_input_tokens is not None
-        else 0
-    )
-    cache_creation_input_tokens = (
-        usage_data.cache_creation_input_tokens
-        if hasattr(usage_data, "cache_creation_input_tokens")
-        and usage_data.cache_creation_input_tokens is not None
-        else 0
-    )
-    return (
-        prompt_tokens,
-        completion_tokens,
-        cache_read_input_tokens,
-        cache_creation_input_tokens,
-    )
-
-
-def _extract_together_tokens(usage_data) -> tuple[int, int, int, int]:
-    """Extract token counts from Together usage data."""
-    prompt_tokens = (
-        usage_data.prompt_tokens
-        if hasattr(usage_data, "prompt_tokens") and usage_data.prompt_tokens is not None
-        else 0
-    )
-    completion_tokens = (
-        usage_data.completion_tokens
-        if hasattr(usage_data, "completion_tokens")
-        and usage_data.completion_tokens is not None
-        else 0
-    )
-    return prompt_tokens, completion_tokens, 0, 0
-
-
-def _extract_groq_tokens(usage_data) -> tuple[int, int, int, int]:
-    """Extract token counts from Groq usage data."""
-    prompt_tokens = (
-        usage_data.prompt_tokens
-        if hasattr(usage_data, "prompt_tokens") and usage_data.prompt_tokens is not None
-        else 0
-    )
-    completion_tokens = (
-        usage_data.completion_tokens
-        if hasattr(usage_data, "completion_tokens")
-        and usage_data.completion_tokens is not None
-        else 0
-    )
-    # Extract cached tokens from prompt_tokens_details.cached_tokens
-    cache_read_input_tokens = 0
-    if (
-        hasattr(usage_data, "prompt_tokens_details")
-        and usage_data.prompt_tokens_details
-    ):
-        if (
-            hasattr(usage_data.prompt_tokens_details, "cached_tokens")
-            and usage_data.prompt_tokens_details.cached_tokens is not None
-        ):
-            cache_read_input_tokens = usage_data.prompt_tokens_details.cached_tokens
-
-    return prompt_tokens, completion_tokens, cache_read_input_tokens, 0
-
-
-# Provider-specific output formatting handlers
-def _format_openai_output(response: Any) -> tuple[Optional[str], Optional[TraceUsage]]:
-    """Format output data from OpenAI response."""
-    from judgeval.tracer.llm.providers import (
-        openai_ChatCompletion,
-        openai_Response,
-        openai_ParsedChatCompletion,
-    )
-
-    model_name = None
-    message_content = None
-    prompt_tokens = 0
-    completion_tokens = 0
-    cache_read_input_tokens = 0
-    cache_creation_input_tokens = 0
-
-    if openai_ChatCompletion and isinstance(response, openai_ChatCompletion):
-        model_name = response.model or ""
-        prompt_tokens = (
-            response.usage.prompt_tokens
-            if response.usage and response.usage.prompt_tokens is not None
-            else 0
-        )
-        completion_tokens = (
-            response.usage.completion_tokens
-            if response.usage and response.usage.completion_tokens is not None
-            else 0
-        )
-        cache_read_input_tokens = (
-            response.usage.prompt_tokens_details.cached_tokens
-            if response.usage
-            and response.usage.prompt_tokens_details
-            and response.usage.prompt_tokens_details.cached_tokens is not None
-            else 0
-        )
-
-        if openai_ParsedChatCompletion and isinstance(
-            response, openai_ParsedChatCompletion
-        ):
-            message_content = response.choices[0].message.parsed
-        else:
-            message_content = response.choices[0].message.content
-    elif openai_Response and isinstance(response, openai_Response):
-        model_name = response.model or ""
-        prompt_tokens = (
-            response.usage.input_tokens
-            if response.usage and response.usage.input_tokens is not None
-            else 0
-        )
-        completion_tokens = (
-            response.usage.output_tokens
-            if response.usage and response.usage.output_tokens is not None
-            else 0
-        )
-        cache_read_input_tokens = (
-            response.usage.input_tokens_details.cached_tokens
-            if response.usage
-            and response.usage.input_tokens_details
-            and response.usage.input_tokens_details.cached_tokens is not None
-            else 0
-        )
-        output0 = response.output[0]
-        if (
-            hasattr(output0, "content")
-            and output0.content
-            and hasattr(output0.content, "__iter__")
-        ):
-            message_content = "".join(
-                seg.text for seg in output0.content if hasattr(seg, "text") and seg.text
-            )
-
-    if model_name:
-        return message_content, _create_usage(
-            model_name,
-            prompt_tokens,
-            completion_tokens,
-            cache_read_input_tokens,
-            cache_creation_input_tokens,
-        )
-
-    return None, None
-
-
-def _format_anthropic_output(
-    response: Any,
-) -> tuple[Optional[str], Optional[TraceUsage]]:
-    """Format output data from Anthropic response."""
-    model_name = getattr(response, "model", "") or ""
-    usage = getattr(response, "usage", None)
-    prompt_tokens = (
-        usage.input_tokens
-        if usage and hasattr(usage, "input_tokens") and usage.input_tokens is not None
-        else 0
-    )
-    completion_tokens = (
-        usage.output_tokens
-        if usage and hasattr(usage, "output_tokens") and usage.output_tokens is not None
-        else 0
-    )
-    cache_read_input_tokens = (
-        usage.cache_read_input_tokens
-        if usage
-        and hasattr(usage, "cache_read_input_tokens")
-        and usage.cache_read_input_tokens is not None
-        else 0
-    )
-    cache_creation_input_tokens = (
-        usage.cache_creation_input_tokens
-        if usage
-        and hasattr(usage, "cache_creation_input_tokens")
-        and usage.cache_creation_input_tokens is not None
-        else 0
-    )
-    # Extract content from Anthropic response, handling both text and tool use blocks
-    message_content = None
-    if hasattr(response, "content") and response.content:
-        content_parts = []
-        for content_block in response.content:
-            block_type = getattr(content_block, "type", None)
-            if block_type == "text":
-                # Text content block
-                content_parts.append(getattr(content_block, "text", ""))
-            elif block_type == "tool_use":
-                # Tool use block - serialize the tool call information
-                tool_info = {
-                    "type": "tool_use",
-                    "id": getattr(content_block, "id", None),
-                    "name": getattr(content_block, "name", None),
-                    "input": getattr(content_block, "input", None),
-                }
-                content_parts.append(f"[TOOL_USE: {tool_info}]")
-        message_content = "\n".join(content_parts) if content_parts else None
-
-    if model_name:
-        return message_content, _create_usage(
-            model_name,
-            prompt_tokens,
-            completion_tokens,
-            cache_read_input_tokens,
-            cache_creation_input_tokens,
-        )
-
-    return None, None
-
-
-def _format_together_output(
-    response: Any,
-) -> tuple[Optional[str], Optional[TraceUsage]]:
-    """Format output data from Together response."""
-    model_name = (response.model or "") if hasattr(response, "model") else ""
-    prompt_tokens = (
-        response.usage.prompt_tokens
-        if hasattr(response.usage, "prompt_tokens")
-        and response.usage.prompt_tokens is not None
-        else 0
-    )
-    completion_tokens = (
-        response.usage.completion_tokens
-        if hasattr(response.usage, "completion_tokens")
-        and response.usage.completion_tokens is not None
-        else 0
-    )
-    message_content = (
-        response.choices[0].message.content if hasattr(response, "choices") else None
-    )
-
-    if model_name:
-        model_name = "together_ai/" + model_name
-        return message_content, _create_usage(
-            model_name,
-            prompt_tokens,
-            completion_tokens,
-            0,
-            0,
-        )
-
-    return None, None
-
-
-def _format_google_output(response: Any) -> tuple[Optional[str], Optional[TraceUsage]]:
-    """Format output data from Google GenAI response."""
-    model_name = getattr(response, "model_version", "") or ""
-    usage_metadata = getattr(response, "usage_metadata", None)
-    prompt_tokens = (
-        usage_metadata.prompt_token_count
-        if usage_metadata
-        and hasattr(usage_metadata, "prompt_token_count")
-        and usage_metadata.prompt_token_count is not None
-        else 0
-    )
-    completion_tokens = (
-        usage_metadata.candidates_token_count
-        if usage_metadata
-        and hasattr(usage_metadata, "candidates_token_count")
-        and usage_metadata.candidates_token_count is not None
-        else 0
-    )
-    message_content = (
-        response.candidates[0].content.parts[0].text
-        if hasattr(response, "candidates")
-        else None
-    )
-
-    cache_read_input_tokens = 0
-    if usage_metadata and hasattr(usage_metadata, "cached_content_token_count"):
-        cache_read_input_tokens = usage_metadata.cached_content_token_count or 0
-
-    if model_name:
-        return message_content, _create_usage(
-            model_name,
-            prompt_tokens,
-            completion_tokens,
-            cache_read_input_tokens,
-            0,
-        )
-
-    return None, None
-
-
-def _format_groq_output(response: Any) -> tuple[Optional[str], Optional[TraceUsage]]:
-    """Format output data from Groq response."""
-    model_name = (response.model or "") if hasattr(response, "model") else ""
-    prompt_tokens = (
-        response.usage.prompt_tokens
-        if hasattr(response.usage, "prompt_tokens")
-        and response.usage.prompt_tokens is not None
-        else 0
-    )
-    completion_tokens = (
-        response.usage.completion_tokens
-        if hasattr(response.usage, "completion_tokens")
-        and response.usage.completion_tokens is not None
-        else 0
-    )
-    # Extract cached tokens from prompt_tokens_details.cached_tokens
-    cache_read_input_tokens = 0
-    if (
-        hasattr(response, "usage")
-        and response.usage
-        and hasattr(response.usage, "prompt_tokens_details")
-        and response.usage.prompt_tokens_details
-    ):
-        if (
-            hasattr(response.usage.prompt_tokens_details, "cached_tokens")
-            and response.usage.prompt_tokens_details.cached_tokens is not None
-        ):
-            cache_read_input_tokens = response.usage.prompt_tokens_details.cached_tokens
-
-    message_content = (
-        response.choices[0].message.content if hasattr(response, "choices") else None
-    )
-
-    if model_name:
-        model_name = "groq/" + model_name
-        return message_content, _create_usage(
-            model_name,
-            prompt_tokens,
-            completion_tokens,
-            cache_read_input_tokens,
-            0,
-        )
-
-    return None, None
-
-
-class _TracedGeneratorBase:
-    """Base class with common logic for parsing stream chunks."""
-
-    __slots__ = (
-        "tracer",
-        "client",
-        "span",
-        "accumulated_content",
-        "model_name",
-        "provider_type",
-    )
-
-    tracer: Tracer
-    client: ApiClient
-    span: Span
-    accumulated_content: str
-    model_name: str
-    provider_type: ProviderType
-
-    def __init__(self, tracer: Tracer, client: ApiClient, span: Span, model_name: str):
-        """Initialize the base traced generator.
-
-        Args:
-            tracer: The tracer instance
-            client: The API client
-            span: The OpenTelemetry span
-            model_name: The model name (empty string default allows fallback to usage_data.model)
-        """
-        self.tracer = tracer
-        self.client = client
-        self.span = span
-        self.accumulated_content = ""
-        self.model_name = model_name
-        self.provider_type = _detect_provider(client)
-
-    def _extract_content(self, chunk) -> str:
-        """Extract content from streaming chunk based on provider."""
-        if self.provider_type == ProviderType.OPENAI:
-            return _extract_openai_content(chunk)
-        elif self.provider_type == ProviderType.ANTHROPIC:
-            return _extract_anthropic_content(chunk)
-        elif self.provider_type == ProviderType.TOGETHER:
-            return _extract_together_content(chunk)
-        elif self.provider_type == ProviderType.GROQ:
-            return _extract_groq_content(chunk)
-        else:
-            # Default case - assume OpenAI-compatible for unknown providers
-            return _extract_openai_content(chunk)
-
-    def _process_chunk_usage(self, chunk):
-        """Process usage data from streaming chunks based on provider."""
-        usage_data = _extract_chunk_usage(self.client, chunk)
-        if usage_data:
-            _process_usage_data(
-                self.span, usage_data, self.tracer, self.client, self.model_name
-            )
-
-    def __del__(self):
-        """
-        Fallback cleanup for unclosed spans. This is a safety mechanism only - spans
-        should normally be finalized in StopIteration/StopAsyncIteration handlers.
-
-        Note: __del__ is not guaranteed to be called in all situations (e.g., reference
-        cycles, program exit), so this should not be relied upon as the primary cleanup
-        mechanism. The primary finalization happens in the iterator protocol methods.
-        """
-        if self.span:
-            try:
-                self._finalize_span()
-            except Exception as e:
-                judgeval_logger.warning(
-                    f"Error during span finalization in __del__: {e}"
-                )
-
-    def _finalize_span(self):
-        """Finalize the span by setting completion content and ending it."""
-        if self.span:
-            set_span_attribute(
-                self.span, AttributeKeys.GEN_AI_COMPLETION, self.accumulated_content
-            )
-            self.span.end()
-            self.span = None
-
-
-class TracedGenerator(_TracedGeneratorBase):
-    """Generator wrapper that adds OpenTelemetry tracing without consuming the stream."""
-
-    __slots__ = ("generator",)
-
-    generator: Union[Generator[Any, None, None], Iterator[Any]]
-
-    def __init__(
-        self,
-        tracer: Tracer,
-        generator: Union[Generator[Any, None, None], Iterator[Any]],
-        client: ApiClient,
-        span: Span,
-        model_name: str,
-    ):
-        super().__init__(tracer, client, span, model_name)
-        self.generator = generator
-
-    def __iter__(self):
-        return self
-
-    def __next__(self):
-        try:
-            chunk = next(self.generator)
-
-            content = self._extract_content(chunk)
-            if content:
-                self.accumulated_content += content
-            self._process_chunk_usage(chunk)
-
-            return chunk
-
-        except StopIteration:
-            self._finalize_span()
-            raise
-        except Exception as e:
-            if self.span:
-                self.span.record_exception(e)
-                self.span.end()
-            raise
-
-
-class TracedAsyncGenerator(_TracedGeneratorBase):
-    """Async generator wrapper that adds OpenTelemetry tracing without consuming the stream."""
-
-    __slots__ = ("async_generator",)
-
-    async_generator: Union[AsyncGenerator[Any, None], AsyncIterator[Any]]
-
-    def __init__(
-        self,
-        tracer: Tracer,
-        async_generator: Union[AsyncGenerator[Any, None], AsyncIterator[Any]],
-        client: ApiClient,
-        span: Span,
-        model_name: str,
-    ):
-        super().__init__(tracer, client, span, model_name)
-        self.async_generator = async_generator
-
-    def __aiter__(self):
-        return self
-
-    async def __anext__(self):
-        try:
-            chunk = await self.async_generator.__anext__()
-
-            content = self._extract_content(chunk)
-            if content:
-                self.accumulated_content += content
-
-            self._process_chunk_usage(chunk)
-
-            return chunk
-
-        except StopAsyncIteration:
-            self._finalize_span()
-            raise
-        except Exception as e:
-            if self.span:
-                self.span.record_exception(e)
-                self.span.end()
-            raise
-
-
-class TracedSyncContextManager:
-    """Sync context manager wrapper for streaming methods."""
-
-    def __init__(
-        self,
-        tracer: Tracer,
-        context_manager: Any,
-        client: ApiClient,
-        span: Span,
-        model_name: str,
-    ):
-        self.tracer = tracer
-        self.context_manager = context_manager
-        self.client = client
-        self.span = span
-        self.stream: Optional[Any] = None
-        self.model_name = model_name
-
-    def __enter__(self):
-        self.stream = self.context_manager.__enter__()
-        return TracedGenerator(
-            self.tracer, self.stream, self.client, self.span, self.model_name
-        )
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        return self.context_manager.__exit__(exc_type, exc_val, exc_tb)
-
-    def __del__(self):
-        """Cleanup span if not properly closed."""
-        if self.span:
-            try:
-                self.span.end()
-            except Exception:
-                pass
-
-
-class TracedAsyncContextManager:
-    """Async context manager wrapper for streaming methods."""
-
-    def __init__(
-        self,
-        tracer: Tracer,
-        context_manager: Any,
-        client: ApiClient,
-        span: Span,
-        model_name: str,
-    ):
-        self.tracer = tracer
-        self.context_manager = context_manager
-        self.client = client
-        self.span = span
-        self.stream: Optional[Any] = None
-        self.model_name = model_name
-
-    async def __aenter__(self):
-        self.stream = await self.context_manager.__aenter__()
-        return TracedAsyncGenerator(
-            self.tracer, self.stream, self.client, self.span, self.model_name
-        )
-
-    async def __aexit__(self, exc_type, exc_val, exc_tb):
-        return await self.context_manager.__aexit__(exc_type, exc_val, exc_tb)
-
-    def __del__(self):
-        """Cleanup span if not properly closed."""
-        if self.span:
-            try:
-                self.span.end()
-            except Exception:
-                pass
-
-
-def _extract_chunk_usage(client: ApiClient, chunk) -> Any:
-    """Extract usage data from streaming chunks based on provider."""
-    provider_type = _detect_provider(client)
-
-    if provider_type == ProviderType.OPENAI:
-        return _extract_openai_chunk_usage(chunk)
-    elif provider_type == ProviderType.ANTHROPIC:
-        return _extract_anthropic_chunk_usage(chunk)
-    elif provider_type == ProviderType.TOGETHER:
-        return _extract_together_chunk_usage(chunk)
-    elif provider_type == ProviderType.GROQ:
-        return _extract_groq_chunk_usage(chunk)
-    else:
-        # Default case - assume OpenAI-compatible for unknown providers
-        return _extract_openai_chunk_usage(chunk)
-
-
-def _extract_usage_tokens(client: ApiClient, usage_data) -> tuple[int, int, int, int]:
-    """Extract token counts from usage data based on provider."""
-    provider_type = _detect_provider(client)
-
-    if provider_type == ProviderType.OPENAI:
-        return _extract_openai_tokens(usage_data)
-    elif provider_type == ProviderType.ANTHROPIC:
-        return _extract_anthropic_tokens(usage_data)
-    elif provider_type == ProviderType.TOGETHER:
-        return _extract_together_tokens(usage_data)
-    elif provider_type == ProviderType.GROQ:
-        return _extract_groq_tokens(usage_data)
-    else:
-        # Default case - assume OpenAI-compatible for unknown providers
-        return _extract_openai_tokens(usage_data)
-
-
-def _process_usage_data(
-    span, usage_data, tracer: Tracer, client: ApiClient, model_name: str
-):
-    """Process usage data and set span attributes."""
-    (
-        prompt_tokens,
-        completion_tokens,
-        cache_read_input_tokens,
-        cache_creation_input_tokens,
-    ) = _extract_usage_tokens(client, usage_data)
-
-    if prompt_tokens or completion_tokens:
-        final_model_name = getattr(usage_data, "model", None) or model_name
-
-        # Add provider prefixes for cost calculation
-        provider_type = _detect_provider(client)
-        if (
-            provider_type == ProviderType.TOGETHER
-            and final_model_name
-            and not final_model_name.startswith("together_ai/")
-        ):
-            final_model_name = "together_ai/" + final_model_name
-        elif (
-            provider_type == ProviderType.GROQ
-            and final_model_name
-            and not final_model_name.startswith("groq/")
-        ):
-            final_model_name = "groq/" + final_model_name
-
-        usage = _create_usage(
-            final_model_name,
-            prompt_tokens,
-            completion_tokens,
-            cache_read_input_tokens,
-            cache_creation_input_tokens,
-        )
-        _set_usage_attributes(span, usage, tracer)
-
-
-def _set_usage_attributes(span, usage: TraceUsage, tracer: Tracer):
-    """Set usage attributes on the span for non-streaming responses."""
-
-    set_span_attribute(span, AttributeKeys.GEN_AI_RESPONSE_MODEL, usage.model_name)
-    set_span_attribute(
-        span, AttributeKeys.GEN_AI_USAGE_INPUT_TOKENS, usage.prompt_tokens
-    )
-    set_span_attribute(
-        span, AttributeKeys.GEN_AI_USAGE_OUTPUT_TOKENS, usage.completion_tokens
-    )
-    set_span_attribute(
-        span, AttributeKeys.GEN_AI_USAGE_COMPLETION_TOKENS, usage.completion_tokens
-    )
-    set_span_attribute(
-        span, AttributeKeys.GEN_AI_USAGE_TOTAL_COST, usage.total_cost_usd
-    )
-
-
-def wrap_provider(tracer: Tracer, client: ApiClient) -> ApiClient:
-    """
-    Wraps an API client to add tracing capabilities.
-    Supports OpenAI, Together, Anthropic, Google GenAI, and Groq clients.
-    """
-
-    def wrapped(function, span_name):
-        @functools.wraps(function)
-        def wrapper(*args, **kwargs):
-            if kwargs.get("stream", False):
-                span = tracer.get_tracer().start_span(
-                    span_name, attributes={AttributeKeys.JUDGMENT_SPAN_KIND: "llm"}
-                )
-                tracer.add_agent_attributes_to_span(span)
-                set_span_attribute(
-                    span, AttributeKeys.GEN_AI_PROMPT, safe_serialize(kwargs)
-                )
-                model_name = kwargs.get("model", "")
-
-                # Add provider prefix for Groq clients
-                if HAS_GROQ:
-                    from judgeval.tracer.llm.providers import groq_Groq, groq_AsyncGroq
-
-                    if (
-                        isinstance(client, (groq_Groq, groq_AsyncGroq))
-                        and model_name
-                        and not model_name.startswith("groq/")
-                    ):
-                        model_name = "groq/" + model_name
-
-                response = function(*args, **kwargs)
-                return TracedGenerator(tracer, response, client, span, model_name)
-            else:
-                with sync_span_context(
-                    tracer, span_name, {AttributeKeys.JUDGMENT_SPAN_KIND: "llm"}
-                ) as span:
-                    tracer.add_agent_attributes_to_span(span)
-                    set_span_attribute(
-                        span, AttributeKeys.GEN_AI_PROMPT, safe_serialize(kwargs)
-                    )
-                    try:
-                        response = function(*args, **kwargs)
-                        output, usage = _format_output_data(client, response)
-                        set_span_attribute(
-                            span, AttributeKeys.GEN_AI_COMPLETION, output
-                        )
-                        if usage:
-                            _set_usage_attributes(span, usage, tracer)
-                        return response
-                    except Exception as e:
-                        span.record_exception(e)
-                        raise
-
-        return wrapper
-
-    def wrapped_async(function, span_name):
-        @functools.wraps(function)
-        async def wrapper(*args, **kwargs):
-            if kwargs.get("stream", False):
-                span = tracer.get_tracer().start_span(
-                    span_name, attributes={AttributeKeys.JUDGMENT_SPAN_KIND: "llm"}
-                )
-                tracer.add_agent_attributes_to_span(span)
-                set_span_attribute(
-                    span, AttributeKeys.GEN_AI_PROMPT, safe_serialize(kwargs)
-                )
-                model_name = kwargs.get("model", "")
-
-                # Add provider prefix for Groq clients
-                if HAS_GROQ:
-                    from judgeval.tracer.llm.providers import groq_Groq, groq_AsyncGroq
-
-                    if (
-                        isinstance(client, (groq_Groq, groq_AsyncGroq))
-                        and model_name
-                        and not model_name.startswith("groq/")
-                    ):
-                        model_name = "groq/" + model_name
-
-                response = await function(*args, **kwargs)
-                return TracedAsyncGenerator(tracer, response, client, span, model_name)
-            else:
-                async with async_span_context(
-                    tracer, span_name, {AttributeKeys.JUDGMENT_SPAN_KIND: "llm"}
-                ) as span:
-                    tracer.add_agent_attributes_to_span(span)
-                    set_span_attribute(
-                        span, AttributeKeys.GEN_AI_PROMPT, safe_serialize(kwargs)
-                    )
-                    try:
-                        response = await function(*args, **kwargs)
-                        output, usage = _format_output_data(client, response)
-                        set_span_attribute(
-                            span, AttributeKeys.GEN_AI_COMPLETION, output
-                        )
-                        if usage:
-                            _set_usage_attributes(span, usage, tracer)
-                        return response
-                    except Exception as e:
-                        span.record_exception(e)
-                        raise
-
-        return wrapper
-
-    def wrapped_sync_context_manager(function, span_name):
-        """Special wrapper for sync context manager methods."""
-
-        @functools.wraps(function)
-        def wrapper(*args, **kwargs):
-            span = tracer.get_tracer().start_span(
-                span_name, attributes={AttributeKeys.JUDGMENT_SPAN_KIND: "llm"}
-            )
-            tracer.add_agent_attributes_to_span(span)
-            set_span_attribute(
-                span, AttributeKeys.GEN_AI_PROMPT, safe_serialize(kwargs)
-            )
-            model_name = kwargs.get("model", "")
-
-            # Add provider prefix for Groq clients
-            if HAS_GROQ:
-                from judgeval.tracer.llm.providers import groq_Groq, groq_AsyncGroq
-
-                if (
-                    isinstance(client, (groq_Groq, groq_AsyncGroq))
-                    and model_name
-                    and not model_name.startswith("groq/")
-                ):
-                    model_name = "groq/" + model_name
-
-            original_context_manager = function(*args, **kwargs)
-            return TracedSyncContextManager(
-                tracer, original_context_manager, client, span, model_name
-            )
-
-        return wrapper
-
-    def wrapped_async_context_manager(function, span_name):
-        """Special wrapper for async context manager methods."""
-
-        @functools.wraps(function)
-        def wrapper(*args, **kwargs):
-            span = tracer.get_tracer().start_span(
-                span_name, attributes={AttributeKeys.JUDGMENT_SPAN_KIND: "llm"}
-            )
-            tracer.add_agent_attributes_to_span(span)
-            set_span_attribute(
-                span, AttributeKeys.GEN_AI_PROMPT, safe_serialize(kwargs)
-            )
-            model_name = kwargs.get("model", "")
-
-            # Add provider prefix for Groq clients
-            if HAS_GROQ:
-                from judgeval.tracer.llm.providers import groq_Groq, groq_AsyncGroq
-
-                if (
-                    isinstance(client, (groq_Groq, groq_AsyncGroq))
-                    and model_name
-                    and not model_name.startswith("groq/")
-                ):
-                    model_name = "groq/" + model_name
-
-            original_context_manager = function(*args, **kwargs)
-            return TracedAsyncContextManager(
-                tracer, original_context_manager, client, span, model_name
-            )
-
-        return wrapper
-
-    if HAS_OPENAI:
-        from judgeval.tracer.llm.providers import openai_OpenAI, openai_AsyncOpenAI
-
-        assert openai_OpenAI is not None, "OpenAI client not found"
-        assert openai_AsyncOpenAI is not None, "OpenAI async client not found"
-        span_name = "OPENAI_API_CALL"
-        if isinstance(client, openai_OpenAI):
-            setattr(
-                client.chat.completions,
-                "create",
-                wrapped(client.chat.completions.create, span_name),
-            )
-            setattr(
-                client.responses, "create", wrapped(client.responses.create, span_name)
-            )
-            setattr(
-                client.beta.chat.completions,
-                "parse",
-                wrapped(client.beta.chat.completions.parse, span_name),
-            )
-        elif isinstance(client, openai_AsyncOpenAI):
-            setattr(
-                client.chat.completions,
-                "create",
-                wrapped_async(client.chat.completions.create, span_name),
-            )
-            setattr(
-                client.responses,
-                "create",
-                wrapped_async(client.responses.create, span_name),
-            )
-            setattr(
-                client.beta.chat.completions,
-                "parse",
-                wrapped_async(client.beta.chat.completions.parse, span_name),
-            )
-
-    if HAS_TOGETHER:
-        from judgeval.tracer.llm.providers import (
-            together_Together,
-            together_AsyncTogether,
-        )
-
-        assert together_Together is not None, "Together client not found"
-        assert together_AsyncTogether is not None, "Together async client not found"
-        span_name = "TOGETHER_API_CALL"
-        if isinstance(client, together_Together):
-            setattr(
-                client.chat.completions,
-                "create",
-                wrapped(client.chat.completions.create, span_name),
-            )
-        elif isinstance(client, together_AsyncTogether):
-            setattr(
-                client.chat.completions,
-                "create",
-                wrapped_async(client.chat.completions.create, span_name),
-            )
-
-    if HAS_ANTHROPIC:
-        from judgeval.tracer.llm.providers import (
-            anthropic_Anthropic,
-            anthropic_AsyncAnthropic,
-        )
-
-        assert anthropic_Anthropic is not None, "Anthropic client not found"
-        assert anthropic_AsyncAnthropic is not None, "Anthropic async client not found"
-        span_name = "ANTHROPIC_API_CALL"
-        if isinstance(client, anthropic_Anthropic):
-            setattr(
-                client.messages, "create", wrapped(client.messages.create, span_name)
-            )
-            setattr(
-                client.messages,
-                "stream",
-                wrapped_sync_context_manager(client.messages.stream, span_name),
-            )
-        elif isinstance(client, anthropic_AsyncAnthropic):
-            setattr(
-                client.messages,
-                "create",
-                wrapped_async(client.messages.create, span_name),
-            )
-            setattr(
-                client.messages,
-                "stream",
-                wrapped_async_context_manager(client.messages.stream, span_name),
-            )
-
-    if HAS_GOOGLE_GENAI:
-        from judgeval.tracer.llm.providers import (
-            google_genai_Client,
-            google_genai_AsyncClient,
-        )
-
-        assert google_genai_Client is not None, "Google GenAI client not found"
-        assert google_genai_AsyncClient is not None, (
-            "Google GenAI async client not found"
-        )
-        span_name = "GOOGLE_API_CALL"
-        if isinstance(client, google_genai_Client):
-            setattr(
-                client.models,
-                "generate_content",
-                wrapped(client.models.generate_content, span_name),
-            )
-        elif isinstance(client, google_genai_AsyncClient):
-            setattr(
-                client.models,
-                "generate_content",
-                wrapped_async(client.models.generate_content, span_name),
-            )
-
-    if HAS_GROQ:
-        from judgeval.tracer.llm.providers import groq_Groq, groq_AsyncGroq
-
-        assert groq_Groq is not None, "Groq client not found"
-        assert groq_AsyncGroq is not None, "Groq async client not found"
-        span_name = "GROQ_API_CALL"
-        if isinstance(client, groq_Groq):
-            setattr(
-                client.chat.completions,
-                "create",
-                wrapped(client.chat.completions.create, span_name),
-            )
-        elif isinstance(client, groq_AsyncGroq):
-            setattr(
-                client.chat.completions,
-                "create",
-                wrapped_async(client.chat.completions.create, span_name),
-            )
-
-    return client
-
-
-def _format_output_data(
-    client: ApiClient, response: Any
-) -> tuple[Optional[str], Optional[TraceUsage]]:
-    """Format output data from LLM response based on provider."""
-    provider_type = _detect_provider(client)
-
-    if provider_type == ProviderType.OPENAI:
-        return _format_openai_output(response)
-    elif provider_type == ProviderType.ANTHROPIC:
-        return _format_anthropic_output(response)
-    elif provider_type == ProviderType.TOGETHER:
-        return _format_together_output(response)
-    elif provider_type == ProviderType.GOOGLE:
-        return _format_google_output(response)
-    elif provider_type == ProviderType.GROQ:
-        return _format_groq_output(response)
-    else:
-        # Default case - assume OpenAI-compatible for unknown providers
-        judgeval_logger.info(
-            f"Unknown client type {type(client)}, assuming OpenAI-compatible"
-        )
-        return _format_openai_output(response)
-
-
-def _create_usage(
-    model_name: str,
-    prompt_tokens: int,
-    completion_tokens: int,
-    cache_read_input_tokens: int = 0,
-    cache_creation_input_tokens: int = 0,
-) -> TraceUsage:
-    prompt_cost, completion_cost = cost_per_token(
-        model=model_name,
-        prompt_tokens=prompt_tokens,
-        completion_tokens=completion_tokens,
-        cache_read_input_tokens=cache_read_input_tokens,
-        cache_creation_input_tokens=cache_creation_input_tokens,
-    )
-    total_cost_usd = (
-        (prompt_cost + completion_cost) if prompt_cost and completion_cost else None
-    )
-    return TraceUsage(
-        prompt_tokens=prompt_tokens,
-        completion_tokens=completion_tokens,
-        total_tokens=prompt_tokens + completion_tokens,
-        cache_read_input_tokens=cache_read_input_tokens,
-        cache_creation_input_tokens=cache_creation_input_tokens,
-        prompt_tokens_cost_usd=prompt_cost,
-        completion_tokens_cost_usd=completion_cost,
-        total_cost_usd=total_cost_usd,
-        model_name=model_name,
-    )
+__all__ = ["_detect_provider", "wrap_provider"]