judgeval 0.1.0__py3-none-any.whl → 0.23.0__py3-none-any.whl

judgeval/v1/instrumentation/llm/llm_together/chat_completions.py

@@ -0,0 +1,382 @@
+from __future__ import annotations
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Awaitable,
+    Callable,
+    Dict,
+    Iterator,
+    AsyncIterator,
+    Generator,
+    AsyncGenerator,
+)
+
+from opentelemetry.trace import Status, StatusCode
+from judgeval.judgment_attribute_keys import AttributeKeys
+from judgeval.utils.serialize import safe_serialize
+from judgeval.utils.wrappers import (
+    immutable_wrap_async,
+    immutable_wrap_sync,
+    mutable_wrap_sync,
+    mutable_wrap_async,
+    immutable_wrap_sync_iterator,
+    immutable_wrap_async_iterator,
+)
+
+if TYPE_CHECKING:
+    from judgeval.v1.tracer import BaseTracer
+    from together import Together, AsyncTogether  # type: ignore[import-untyped]
+    from together.types import ChatCompletionResponse, ChatCompletionChunk  # type: ignore[import-untyped]
+    from together.types.common import UsageData  # type: ignore[import-untyped]
+
+
+def _extract_together_tokens(usage: UsageData) -> tuple[int, int, int, int]:
+    prompt_tokens = usage.prompt_tokens if usage.prompt_tokens is not None else 0
+    completion_tokens = (
+        usage.completion_tokens if usage.completion_tokens is not None else 0
+    )
+    cache_read_input_tokens = 0
+    cache_creation_input_tokens = 0
+    return (
+        prompt_tokens,
+        completion_tokens,
+        cache_read_input_tokens,
+        cache_creation_input_tokens,
+    )
+
+
+def wrap_chat_completions_create_sync(tracer: BaseTracer, client: Together) -> None:
+    original_func = client.chat.completions.create
+
+    def dispatcher(*args: Any, **kwargs: Any) -> Any:
+        if kwargs.get("stream", False):
+            return _wrap_streaming_sync(tracer, original_func)(*args, **kwargs)  # type: ignore[arg-type]
+        return _wrap_non_streaming_sync(tracer, original_func)(*args, **kwargs)  # type: ignore[arg-type]
+
+    setattr(client.chat.completions, "create", dispatcher)
+
+
+def _wrap_non_streaming_sync(
+    tracer: BaseTracer, original_func: Callable[..., ChatCompletionResponse]
+) -> Callable[..., ChatCompletionResponse]:
+    def pre_hook(ctx: Dict[str, Any], *args: Any, **kwargs: Any) -> None:
+        ctx["span"] = tracer.get_tracer().start_span(
+            "TOGETHER_API_CALL", attributes={AttributeKeys.JUDGMENT_SPAN_KIND: "llm"}
+        )
+        ctx["span"].set_attribute(AttributeKeys.GEN_AI_PROMPT, safe_serialize(kwargs))
+        ctx["model_name"] = kwargs.get("model", "")
+        prefixed_model_name = (
+            f"together_ai/{ctx['model_name']}" if ctx["model_name"] else ""
+        )
+        ctx["model_name"] = prefixed_model_name
+        ctx["span"].set_attribute(
+            AttributeKeys.JUDGMENT_LLM_MODEL_NAME, prefixed_model_name
+        )
+
+    def post_hook(ctx: Dict[str, Any], result: ChatCompletionResponse) -> None:
+        span = ctx.get("span")
+        if not span:
+            return
+
+        span.set_attribute(AttributeKeys.GEN_AI_COMPLETION, safe_serialize(result))
+
+        if result.usage:
+            prompt_tokens, completion_tokens, _, _ = _extract_together_tokens(
+                result.usage
+            )
+            span.set_attribute(
+                AttributeKeys.JUDGMENT_USAGE_NON_CACHED_INPUT_TOKENS,
+                prompt_tokens,
+            )
+            span.set_attribute(
+                AttributeKeys.JUDGMENT_USAGE_OUTPUT_TOKENS, completion_tokens
+            )
+            span.set_attribute(
+                AttributeKeys.JUDGMENT_USAGE_METADATA,
+                safe_serialize(result.usage),
+            )
+
+        span.set_attribute(AttributeKeys.JUDGMENT_LLM_MODEL_NAME, ctx["model_name"])
+
+    def error_hook(ctx: Dict[str, Any], error: Exception) -> None:
+        span = ctx.get("span")
+        if span:
+            span.record_exception(error)
+            span.set_status(Status(StatusCode.ERROR))
+
+    def finally_hook(ctx: Dict[str, Any]) -> None:
+        span = ctx.get("span")
+        if span:
+            span.end()
+
+    return immutable_wrap_sync(
+        original_func,
+        pre_hook=pre_hook,
+        post_hook=post_hook,
+        error_hook=error_hook,
+        finally_hook=finally_hook,
+    )
+
+
+def _wrap_streaming_sync(
+    tracer: BaseTracer, original_func: Callable[..., Iterator[ChatCompletionChunk]]
+) -> Callable[..., Iterator[ChatCompletionChunk]]:
+    def pre_hook(ctx: Dict[str, Any], *args: Any, **kwargs: Any) -> None:
+        ctx["span"] = tracer.get_tracer().start_span(
+            "TOGETHER_API_CALL", attributes={AttributeKeys.JUDGMENT_SPAN_KIND: "llm"}
+        )
+        ctx["span"].set_attribute(AttributeKeys.GEN_AI_PROMPT, safe_serialize(kwargs))
+        ctx["model_name"] = kwargs.get("model", "")
+        prefixed_model_name = (
+            f"together_ai/{ctx['model_name']}" if ctx["model_name"] else ""
+        )
+        ctx["model_name"] = prefixed_model_name
+        ctx["span"].set_attribute(
+            AttributeKeys.JUDGMENT_LLM_MODEL_NAME, prefixed_model_name
+        )
+        ctx["accumulated_content"] = ""
+
+    def mutate_hook(
+        ctx: Dict[str, Any], result: Iterator[ChatCompletionChunk]
+    ) -> Iterator[ChatCompletionChunk]:
+        def traced_generator() -> Generator[ChatCompletionChunk, None, None]:
+            for chunk in result:
+                yield chunk
+
+        def yield_hook(inner_ctx: Dict[str, Any], chunk: ChatCompletionChunk) -> None:
+            span = ctx.get("span")
+            if not span:
+                return
+
+            if chunk.choices and len(chunk.choices) > 0:
+                delta = chunk.choices[0].delta
+                if delta and hasattr(delta, "content") and delta.content:
+                    ctx["accumulated_content"] = (
+                        ctx.get("accumulated_content", "") + delta.content
+                    )
+
+            if chunk.usage:
+                prompt_tokens, completion_tokens, _, _ = _extract_together_tokens(
+                    chunk.usage
+                )
+                span.set_attribute(
+                    AttributeKeys.JUDGMENT_USAGE_NON_CACHED_INPUT_TOKENS,
+                    prompt_tokens,
+                )
+                span.set_attribute(
+                    AttributeKeys.JUDGMENT_USAGE_OUTPUT_TOKENS, completion_tokens
+                )
+                span.set_attribute(
+                    AttributeKeys.JUDGMENT_USAGE_METADATA,
+                    safe_serialize(chunk.usage),
+                )
+
+        def post_hook_inner(inner_ctx: Dict[str, Any]) -> None:
+            span = ctx.get("span")
+            if span:
+                accumulated = ctx.get("accumulated_content", "")
+                span.set_attribute(AttributeKeys.GEN_AI_COMPLETION, accumulated)
+
+        def error_hook_inner(inner_ctx: Dict[str, Any], error: Exception) -> None:
+            span = ctx.get("span")
+            if span:
+                span.record_exception(error)
+                span.set_status(Status(StatusCode.ERROR))
+
+        def finally_hook_inner(inner_ctx: Dict[str, Any]) -> None:
+            span = ctx.get("span")
+            if span:
+                span.end()
+
+        wrapped_generator = immutable_wrap_sync_iterator(
+            traced_generator,
+            yield_hook=yield_hook,
+            post_hook=post_hook_inner,
+            error_hook=error_hook_inner,
+            finally_hook=finally_hook_inner,
+        )
+
+        return wrapped_generator()
+
+    def error_hook(ctx: Dict[str, Any], error: Exception) -> None:
+        span = ctx.get("span")
+        if span:
+            span.record_exception(error)
+            span.set_status(Status(StatusCode.ERROR))
+
+    return mutable_wrap_sync(
+        original_func,
+        pre_hook=pre_hook,
+        mutate_hook=mutate_hook,
+        error_hook=error_hook,
+    )
+
+
+def wrap_chat_completions_create_async(
+    tracer: BaseTracer, client: AsyncTogether
+) -> None:
+    original_func = client.chat.completions.create
+
+    async def dispatcher(*args: Any, **kwargs: Any) -> Any:
+        if kwargs.get("stream", False):
+            return await _wrap_streaming_async(tracer, original_func)(*args, **kwargs)  # type: ignore[arg-type]
+        return await _wrap_non_streaming_async(tracer, original_func)(*args, **kwargs)  # type: ignore[arg-type]
+
+    setattr(client.chat.completions, "create", dispatcher)
+
+
+def _wrap_non_streaming_async(
+    tracer: BaseTracer, original_func: Callable[..., Awaitable[ChatCompletionResponse]]
+) -> Callable[..., Awaitable[ChatCompletionResponse]]:
+    def pre_hook(ctx: Dict[str, Any], *args: Any, **kwargs: Any) -> None:
+        ctx["span"] = tracer.get_tracer().start_span(
+            "TOGETHER_API_CALL", attributes={AttributeKeys.JUDGMENT_SPAN_KIND: "llm"}
+        )
+        ctx["span"].set_attribute(AttributeKeys.GEN_AI_PROMPT, safe_serialize(kwargs))
+        ctx["model_name"] = kwargs.get("model", "")
+        prefixed_model_name = (
+            f"together_ai/{ctx['model_name']}" if ctx["model_name"] else ""
+        )
+        ctx["model_name"] = prefixed_model_name
+        ctx["span"].set_attribute(
+            AttributeKeys.JUDGMENT_LLM_MODEL_NAME, prefixed_model_name
+        )
+
+    def post_hook(ctx: Dict[str, Any], result: ChatCompletionResponse) -> None:
+        span = ctx.get("span")
+        if not span:
+            return
+
+        span.set_attribute(AttributeKeys.GEN_AI_COMPLETION, safe_serialize(result))
+
+        if result.usage:
+            prompt_tokens, completion_tokens, _, _ = _extract_together_tokens(
+                result.usage
+            )
+            span.set_attribute(
+                AttributeKeys.JUDGMENT_USAGE_NON_CACHED_INPUT_TOKENS,
+                prompt_tokens,
+            )
+            span.set_attribute(
+                AttributeKeys.JUDGMENT_USAGE_OUTPUT_TOKENS, completion_tokens
+            )
+            span.set_attribute(
+                AttributeKeys.JUDGMENT_USAGE_METADATA,
+                safe_serialize(result.usage),
+            )
+
+        span.set_attribute(AttributeKeys.JUDGMENT_LLM_MODEL_NAME, ctx["model_name"])
+
+    def error_hook(ctx: Dict[str, Any], error: Exception) -> None:
+        span = ctx.get("span")
+        if span:
+            span.record_exception(error)
+            span.set_status(Status(StatusCode.ERROR))
+
+    def finally_hook(ctx: Dict[str, Any]) -> None:
+        span = ctx.get("span")
+        if span:
+            span.end()
+
+    return immutable_wrap_async(
+        original_func,
+        pre_hook=pre_hook,
+        post_hook=post_hook,
+        error_hook=error_hook,
+        finally_hook=finally_hook,
+    )
+
+
+def _wrap_streaming_async(
+    tracer: BaseTracer,
+    original_func: Callable[..., Awaitable[AsyncIterator[ChatCompletionChunk]]],
+) -> Callable[..., Awaitable[AsyncIterator[ChatCompletionChunk]]]:
+    def pre_hook(ctx: Dict[str, Any], *args: Any, **kwargs: Any) -> None:
+        ctx["span"] = tracer.get_tracer().start_span(
+            "TOGETHER_API_CALL", attributes={AttributeKeys.JUDGMENT_SPAN_KIND: "llm"}
+        )
+        ctx["span"].set_attribute(AttributeKeys.GEN_AI_PROMPT, safe_serialize(kwargs))
+        ctx["model_name"] = kwargs.get("model", "")
+        prefixed_model_name = (
+            f"together_ai/{ctx['model_name']}" if ctx["model_name"] else ""
+        )
+        ctx["model_name"] = prefixed_model_name
+        ctx["span"].set_attribute(
+            AttributeKeys.JUDGMENT_LLM_MODEL_NAME, prefixed_model_name
+        )
+        ctx["accumulated_content"] = ""
+
+    def mutate_hook(
+        ctx: Dict[str, Any], result: AsyncIterator[ChatCompletionChunk]
+    ) -> AsyncIterator[ChatCompletionChunk]:
+        async def traced_generator() -> AsyncGenerator[ChatCompletionChunk, None]:
+            async for chunk in result:
+                yield chunk
+
+        def yield_hook(inner_ctx: Dict[str, Any], chunk: ChatCompletionChunk) -> None:
+            span = ctx.get("span")
+            if not span:
+                return
+
+            if chunk.choices and len(chunk.choices) > 0:
+                delta = chunk.choices[0].delta
+                if delta and hasattr(delta, "content") and delta.content:
+                    ctx["accumulated_content"] = (
+                        ctx.get("accumulated_content", "") + delta.content
+                    )
+
+            if chunk.usage:
+                prompt_tokens, completion_tokens, _, _ = _extract_together_tokens(
+                    chunk.usage
+                )
+                span.set_attribute(
+                    AttributeKeys.JUDGMENT_USAGE_NON_CACHED_INPUT_TOKENS,
+                    prompt_tokens,
+                )
+                span.set_attribute(
+                    AttributeKeys.JUDGMENT_USAGE_OUTPUT_TOKENS, completion_tokens
+                )
+                span.set_attribute(
+                    AttributeKeys.JUDGMENT_USAGE_METADATA,
+                    safe_serialize(chunk.usage),
+                )
+
+        def post_hook_inner(inner_ctx: Dict[str, Any]) -> None:
+            span = ctx.get("span")
+            if span:
+                accumulated = ctx.get("accumulated_content", "")
+                span.set_attribute(AttributeKeys.GEN_AI_COMPLETION, accumulated)
+
+        def error_hook_inner(inner_ctx: Dict[str, Any], error: Exception) -> None:
+            span = ctx.get("span")
+            if span:
+                span.record_exception(error)
+                span.set_status(Status(StatusCode.ERROR))
+
+        def finally_hook_inner(inner_ctx: Dict[str, Any]) -> None:
+            span = ctx.get("span")
+            if span:
+                span.end()
+
+        wrapped_generator = immutable_wrap_async_iterator(
+            traced_generator,
+            yield_hook=yield_hook,
+            post_hook=post_hook_inner,
+            error_hook=error_hook_inner,
+            finally_hook=finally_hook_inner,
+        )
+
+        return wrapped_generator()
+
+    def error_hook(ctx: Dict[str, Any], error: Exception) -> None:
+        span = ctx.get("span")
+        if span:
+            span.record_exception(error)
+            span.set_status(Status(StatusCode.ERROR))
+
+    return mutable_wrap_async(
+        original_func,
+        pre_hook=pre_hook,
+        mutate_hook=mutate_hook,
+        error_hook=error_hook,
+    )

judgeval/v1/instrumentation/llm/llm_together/config.py

@@ -0,0 +1,6 @@
+from __future__ import annotations
+import importlib.util
+
+HAS_TOGETHER = importlib.util.find_spec("together") is not None
+
+__all__ = ["HAS_TOGETHER"]

judgeval/v1/instrumentation/llm/llm_together/wrapper.py

@@ -0,0 +1,57 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING, Union
+import typing
+
+from judgeval.v1.instrumentation.llm.llm_together.chat_completions import (
+    wrap_chat_completions_create_sync,
+    wrap_chat_completions_create_async,
+)
+
+
+if TYPE_CHECKING:
+    from judgeval.v1.tracer import BaseTracer
+    from together import Together, AsyncTogether  # type: ignore[import-untyped]
+
+    TClient = Union[Together, AsyncTogether]
+
+
+def wrap_together_client_sync(tracer: BaseTracer, client: Together) -> Together:
+    wrap_chat_completions_create_sync(tracer, client)
+    return client
+
+
+def wrap_together_client_async(
+    tracer: BaseTracer, client: AsyncTogether
+) -> AsyncTogether:
+    wrap_chat_completions_create_async(tracer, client)
+    return client
+
+
+@typing.overload
+def wrap_together_client(tracer: BaseTracer, client: Together) -> Together: ...
+@typing.overload
+def wrap_together_client(  # type: ignore[overload-cannot-match]
+    tracer: BaseTracer,
+    client: AsyncTogether,
+) -> AsyncTogether: ...
+
+
+def wrap_together_client(tracer: BaseTracer, client: TClient) -> TClient:
+    from judgeval.v1.instrumentation.llm.llm_together.config import HAS_TOGETHER
+    from judgeval.logger import judgeval_logger
+
+    if not HAS_TOGETHER:
+        judgeval_logger.error(
+            "Cannot wrap Together client: 'together' library not installed. "
+            "Install it with: pip install together"
+        )
+        return client
+
+    from together import Together, AsyncTogether  # type: ignore[import-untyped]
+
+    if isinstance(client, AsyncTogether):
+        return wrap_together_client_async(tracer, client)
+    elif isinstance(client, Together):
+        return wrap_together_client_sync(tracer, client)
+    else:
+        raise TypeError(f"Invalid client type: {type(client)}")

judgeval/v1/instrumentation/llm/providers.py

@@ -0,0 +1,19 @@
+from __future__ import annotations
+from typing import Any, TypeAlias
+
+from judgeval.v1.instrumentation.llm.llm_openai.config import HAS_OPENAI
+from judgeval.v1.instrumentation.llm.llm_together.config import HAS_TOGETHER
+from judgeval.v1.instrumentation.llm.llm_anthropic.config import HAS_ANTHROPIC
+from judgeval.v1.instrumentation.llm.llm_google.config import HAS_GOOGLE_GENAI
+
+# TODO: if we support dependency groups we can have this better type, but during runtime, we do
+# not know which clients an end user might have installed.
+ApiClient: TypeAlias = Any
+
+__all__ = [
+    "ApiClient",
+    "HAS_OPENAI",
+    "HAS_TOGETHER",
+    "HAS_ANTHROPIC",
+    "HAS_GOOGLE_GENAI",
+]

judgeval/v1/integrations/claude_agent_sdk/__init__.py

@@ -0,0 +1,119 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING
+import sys
+from judgeval.logger import judgeval_logger
+
+if TYPE_CHECKING:
+    from judgeval.v1.tracer.base_tracer import BaseTracer
+
+__all__ = ["setup_claude_agent_sdk"]
+
+try:
+    import claude_agent_sdk  # type: ignore
+except ImportError:
+    raise ImportError(
+        "Claude Agent SDK is not installed and required for the claude agent sdk integration. Please install it with `pip install claude-agent-sdk`."
+    )
+
+
+def setup_claude_agent_sdk(
+    tracer: "BaseTracer",
+) -> bool:
+    """
+    Setup Judgeval integration with Claude Agent SDK. Will automatically patch the SDK for automatic tracing.
+
+    Args:
+        tracer: Judgeval Tracer instance
+
+    Returns:
+        bool: True if setup was successful, False otherwise.
+
+    Example:
+        ```python
+        import claude_agent_sdk
+        from judgeval.v1.integrations.claude_agent_sdk import setup_claude_agent_sdk
+
+        tracer = Tracer(project_name="my-project")
+        setup_claude_agent_sdk(tracer=tracer)
+
+        # Now use claude_agent_sdk normally - all calls automatically traced
+        ```
+    """
+    from judgeval.v1.integrations.claude_agent_sdk.wrapper import (
+        _create_client_wrapper_class,
+        _create_tool_wrapper_class,
+        _wrap_tool_factory,
+        _wrap_query_function,
+    )
+
+    try:
+        # Store original classes before patching
+        original_client = (
+            claude_agent_sdk.ClaudeSDKClient
+            if hasattr(claude_agent_sdk, "ClaudeSDKClient")
+            else None
+        )
+        original_tool_class = (
+            claude_agent_sdk.SdkMcpTool
+            if hasattr(claude_agent_sdk, "SdkMcpTool")
+            else None
+        )
+        original_tool_fn = (
+            claude_agent_sdk.tool if hasattr(claude_agent_sdk, "tool") else None
+        )
+        original_query_fn = (
+            claude_agent_sdk.query if hasattr(claude_agent_sdk, "query") else None
+        )
+
+        # Patch ClaudeSDKClient
+        if original_client:
+            wrapped_client = _create_client_wrapper_class(original_client, tracer)
+            claude_agent_sdk.ClaudeSDKClient = wrapped_client  # type: ignore
+
+            # Update all modules that already imported ClaudeSDKClient
+            for module in list(sys.modules.values()):
+                if module and hasattr(module, "ClaudeSDKClient"):
+                    if getattr(module, "ClaudeSDKClient", None) is original_client:
+                        setattr(module, "ClaudeSDKClient", wrapped_client)
+
+        # Patch SdkMcpTool
+        if original_tool_class:
+            wrapped_tool_class = _create_tool_wrapper_class(original_tool_class, tracer)
+            claude_agent_sdk.SdkMcpTool = wrapped_tool_class  # type: ignore
+
+            # Update all modules that already imported SdkMcpTool
+            for module in list(sys.modules.values()):
+                if module and hasattr(module, "SdkMcpTool"):
+                    if getattr(module, "SdkMcpTool", None) is original_tool_class:
+                        setattr(module, "SdkMcpTool", wrapped_tool_class)
+
+        # Patch tool() decorator
+        if original_tool_fn:
+            wrapped_tool_fn = _wrap_tool_factory(original_tool_fn, tracer)
+            claude_agent_sdk.tool = wrapped_tool_fn  # type: ignore
+
+            # Update all modules that already imported tool
+            for module in list(sys.modules.values()):
+                if module and hasattr(module, "tool"):
+                    if getattr(module, "tool", None) is original_tool_fn:
+                        setattr(module, "tool", wrapped_tool_fn)
+
+        # Patch standalone query() function if it exists
+        # Note: The standalone query() uses InternalClient, not ClaudeSDKClient,
+        # so we need to wrap it separately to add tracing
+        if original_query_fn:
+            wrapped_query_fn = _wrap_query_function(original_query_fn, tracer)
+            claude_agent_sdk.query = wrapped_query_fn  # type: ignore
+
+            # Update all modules that already imported query
+            for module in list(sys.modules.values()):
+                if module and hasattr(module, "query"):
+                    if getattr(module, "query", None) is original_query_fn:
+                        setattr(module, "query", wrapped_query_fn)
+
+        judgeval_logger.info("Claude Agent SDK integration setup successful")
+        return True
+
+    except Exception as e:
+        judgeval_logger.error(f"Failed to setup Claude Agent SDK integration: {e}")
+        return False