agenta 0.20.0a4__tar.gz → 0.20.0a6__tar.gz

{agenta-0.20.0a4 → agenta-0.20.0a6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: agenta
-Version: 0.20.0a4
+Version: 0.20.0a6
 Summary: The SDK for agenta is an open-source LLMOps platform.
 Home-page: https://agenta.ai
 Keywords: LLMOps,LLM,evaluation,prompt engineering

{agenta-0.20.0a4 → agenta-0.20.0a6}/agenta/client/backend/types/span.py

@@ -26,7 +26,9 @@ class Span(pydantic.BaseModel):
     spankind: str
     status: SpanStatusCode
     metadata: typing.Dict[str, typing.Any]
+    trace_id: str
     user_id: typing.Optional[str]
+    content: typing.Dict[str, typing.Any]
     children: typing.Optional[typing.List[Span]]
 
     def json(self, **kwargs: typing.Any) -> str:

{agenta-0.20.0a4 → agenta-0.20.0a6}/agenta/sdk/decorators/llm_entrypoint.py

@@ -331,32 +331,31 @@ class entrypoint(BaseDecorator):
             if isinstance(result, Context):
                 save_context(result)
 
-            DEFAULT_KEY = "message"
+            data = result
 
-            if isinstance(result, Dict):
+            # PATCH : if result is not a dict, make it a dict
+            if not isinstance(result, dict):
                 data = result
-
-                # EVENTUALLY THIS PATCH SHOULD BE REMOVED
-                # PATCH: if message in result then only keep message key/value
-                # DEFAULT_KEY = "message"
-
-                if "message" in result.keys():
-                    data = {DEFAULT_KEY: result["message"]}
-                # END OF PATCH
-
-            elif isinstance(result, str):
-                data = {DEFAULT_KEY: result}
-            elif isinstance(result, int) or isinstance(result, float):
-                data = {DEFAULT_KEY: str(result)}
+            else:
+                # PATCH : if result is a legacy dict, clean it up
+                if (
+                    "message" in result.keys()
+                    and "cost" in result.keys()
+                    and "usage" in result.keys()
+                ):
+                    data = {"message": result["message"]}
+            # END OF PATH
 
             if data is None:
-                warning = (
+                data = (
                     "Function executed successfully, but did return None. \n Are you sure you did not forget to return a value?",
                 )
 
-                data = {"message": warning}
+            response = BaseResponse(data=data, trace=trace)
+
+            logging.debug(response)
 
-            return BaseResponse(data=data, trace=trace)
+            return response
 
         except Exception as e:
             self.handle_exception(e)

{agenta-0.20.0a4 → agenta-0.20.0a6}/agenta/sdk/decorators/tracing.py

@@ -65,19 +65,24 @@ class instrument(BaseDecorator):
                 ):
                     result = await func(*args, **kwargs)
 
-                    outputs = result
+                    TRACE_DEFAULT_KEY = "__default__"
 
-                    # EVENTUALLY THIS PATCH SHOULD BE REMOVED
-                    # PATCH : if result is not a dict, make it a dict, in span
-                    DEFAULT_KEY = "message"
+                    outputs = result
 
+                    # PATCH : if result is not a dict, make it a dict
                     if not isinstance(result, dict):
-                        value = result
-
-                        if result.__class__.__module__ != "__builtin__":
-                            value = repr(value)
-
-                        outputs = {DEFAULT_KEY: result}
+                        outputs = {TRACE_DEFAULT_KEY: result}
+                    else:
+                        # PATCH : if result is a legacy dict, clean it up
+                        if (
+                            "message" in result.keys()
+                            and "cost" in result.keys()
+                            and "usage" in result.keys()
+                        ):
+                            outputs = {"message": result["message"]}
+
+                            ag.tracing.store_cost(result["cost"])
+                            ag.tracing.store_usage(result["usage"])
                     # END OF PATH
 
                     ag.tracing.store_outputs(outputs)
@@ -97,19 +102,24 @@ class instrument(BaseDecorator):
                 ):
                     result = func(*args, **kwargs)
 
-                    outputs = result
+                    TRACE_DEFAULT_KEY = "__default__"
 
-                    # EVENTUALLY THIS PATCH SHOULD BE REMOVED
-                    # PATCH : if result is not a dict, make it a dict, in span
-                    DEFAULT_KEY = "message"
+                    outputs = result
 
+                    # PATCH : if result is not a dict, make it a dict
                     if not isinstance(result, dict):
-                        value = result
-
-                        if result.__class__.__module__ != "__builtin__":
-                            value = repr(value)
-
-                        outputs = {DEFAULT_KEY: result}
+                        outputs = {TRACE_DEFAULT_KEY: result}
+                    else:
+                        # PATCH : if result is a legacy dict, clean it up
+                        if (
+                            "message" in result.keys()
+                            and "cost" in result.keys()
+                            and "usage" in result.keys()
+                        ):
+                            outputs = {"message": result["message"]}
+
+                            ag.tracing.store_cost(result["cost"])
+                            ag.tracing.store_usage(result["usage"])
                     # END OF PATH
 
                     ag.tracing.store_outputs(outputs)

{agenta-0.20.0a4 → agenta-0.20.0a6}/agenta/sdk/tracing/callbacks.py

@@ -59,51 +59,30 @@ def litellm_handler():
             )
 
         @debug()
-        def log_stream_event(self, kwargs, response_obj, start_time, end_time):
+        def log_stream_event(self, kwargs, res: ModelResponse, start_time, end_time):
             ag.tracing.set_status(status="OK", span_id=self.span.id)
+            ag.tracing.store_cost(kwargs.get("response_cost"))
+            ag.tracing.store_usage(res.usage.dict() if hasattr(res, "usage") else None)
             ag.tracing.store_outputs(
-                outputs={
-                    "message": kwargs.get(
-                        "complete_streaming_response"
-                    ),  # the complete streamed response (only set if `completion(..stream=True)`)
-                    "usage": (
-                        response_obj.usage.dict()
-                        if hasattr(response_obj, "usage")
-                        else None
-                    ),  # litellm calculates usage
-                    "cost": kwargs.get(
-                        "response_cost"
-                    ),  # litellm calculates response cost
-                },
+                # the complete streamed response (only set if `completion(..stream=True)`
+                outputs={"message": kwargs.get("complete_streaming_response")},
                 span_id=self.span.id,
             )
             ag.tracing.close_span(span_id=self.span.id)
 
         @debug()
-        def log_success_event(
-            self, kwargs, response_obj: ModelResponse, start_time, end_time
-        ):
+        def log_success_event(self, kwargs, res: ModelResponse, start_time, end_time):
             ag.tracing.set_status(status="OK", span_id=self.span.id)
+            ag.tracing.store_cost(kwargs.get("response_cost"))
+            ag.tracing.store_usage(res.usage.dict() if hasattr(res, "usage") else None)
             ag.tracing.store_outputs(
-                outputs={
-                    "message": response_obj.choices[0].message.content,
-                    "usage": (
-                        response_obj.usage.dict()
-                        if hasattr(response_obj, "usage")
-                        else None
-                    ),  # litellm calculates usage
-                    "cost": kwargs.get(
-                        "response_cost"
-                    ),  # litellm calculates response cost
-                },
+                outputs={"message": res.choices[0].message.content},
                 span_id=self.span.id,
             )
             ag.tracing.close_span(span_id=self.span.id)
 
         @debug()
-        def log_failure_event(
-            self, kwargs, response_obj: ModelResponse, start_time, end_time
-        ):
+        def log_failure_event(self, kwargs, res: ModelResponse, start_time, end_time):
             ag.tracing.set_status(status="ERROR", span_id=self.span.id)
             ag.tracing.set_attributes(
                 {
@@ -116,69 +95,45 @@ def litellm_handler():
                 },
                 span_id=self.span.id,
             )
+            ag.tracing.store_cost(kwargs.get("response_cost"))
+            ag.tracing.store_usage(res.usage.dict() if hasattr(res, "usage") else None)
             ag.tracing.store_outputs(
-                outputs={
-                    "message": repr(kwargs["exception"]),  # the Exception raised
-                    "usage": (
-                        response_obj.usage.dict()
-                        if hasattr(response_obj, "usage")
-                        else None
-                    ),  # litellm calculates usage
-                    "cost": kwargs.get(
-                        "response_cost"
-                    ),  # litellm calculates response cost
-                },
+                # the Exception raised
+                outputs={"message": repr(kwargs["exception"])},
                 span_id=self.span.id,
             )
             ag.tracing.close_span(span_id=self.span.id)
 
         @debug()
         async def async_log_stream_event(
-            self, kwargs, response_obj, start_time, end_time
+            self, kwargs, res: ModelResponse, start_time, end_time
         ):
             ag.tracing.set_status(status="OK", span_id=self.span.id)
+            ag.tracing.store_cost(kwargs.get("response_cost"))
+            ag.tracing.store_usage(res.usage.dict() if hasattr(res, "usage") else None)
             ag.tracing.store_outputs(
-                outputs={
-                    "message": kwargs.get(
-                        "complete_streaming_response"
-                    ),  # the complete streamed response (only set if `completion(..stream=True)`)
-                    "usage": (
-                        response_obj.usage.dict()
-                        if hasattr(response_obj, "usage")
-                        else None
-                    ),  # litellm calculates usage
-                    "cost": kwargs.get(
-                        "response_cost"
-                    ),  # litellm calculates response cost
-                },
+                # the complete streamed response (only set if `completion(..stream=True)`)
+                outputs={"message": kwargs.get("complete_streaming_response")},
                 span_id=self.span.id,
             )
             ag.tracing.close_span(span_id=self.span.id)
 
         @debug()
         async def async_log_success_event(
-            self, kwargs, response_obj, start_time, end_time
+            self, kwargs, res: ModelResponse, start_time, end_time
         ):
             ag.tracing.set_status(status="OK", span_id=self.span.id)
+            ag.tracing.store_cost(kwargs.get("response_cost"))
+            ag.tracing.store_usage(res.usage.dict() if hasattr(res, "usage") else None)
             ag.tracing.store_outputs(
-                outputs={
-                    "message": response_obj.choices[0].message.content,
-                    "usage": (
-                        response_obj.usage.dict()
-                        if hasattr(response_obj, "usage")
-                        else None
-                    ),  # litellm calculates usage
-                    "cost": kwargs.get(
-                        "response_cost"
-                    ),  # litellm calculates response cost
-                },
+                outputs={"message": res.choices[0].message.content},
                 span_id=self.span.id,
             )
             ag.tracing.close_span(span_id=self.span.id)
 
         @debug()
         async def async_log_failure_event(
-            self, kwargs, response_obj, start_time, end_time
+            self, kwargs, res: ModelResponse, start_time, end_time
         ):
             ag.tracing.set_status(status="ERROR", span_id=self.span.id)
             ag.tracing.set_attributes(
@@ -192,18 +147,11 @@ def litellm_handler():
                 },
                 span_id=self.span.id,
             )
+            ag.tracing.store_cost(kwargs.get("response_cost"))
+            ag.tracing.store_usage(res.usage.dict() if hasattr(res, "usage") else None)
             ag.tracing.store_outputs(
-                outputs={
-                    "message": repr(kwargs["exception"]),  # the Exception raised
-                    "usage": (
-                        response_obj.usage.dict()
-                        if hasattr(response_obj, "usage")
-                        else None
-                    ),  # litellm calculates usage
-                    "cost": kwargs.get(
-                        "response_cost"
-                    ),  # litellm calculates response cost
-                },
+                # the Exception raised
+                outputs={"message": repr(kwargs["exception"])},
                 span_id=self.span.id,
             )
             ag.tracing.close_span(span_id=self.span.id)

{agenta-0.20.0a4 → agenta-0.20.0a6}/agenta/sdk/tracing/llm_tracing.py

@@ -414,6 +414,30 @@ class Tracing(metaclass=SingletonMeta):
 
         span.outputs = outputs
 
+    @debug()
+    def store_cost(self, cost: float = 0.0, span_id: Optional[str] = None) -> None:
+        """
+        ...
+        """
+        span = self._get_target_span(span_id)
+
+        logging.info(f"Setting span  {span.id} {span.spankind.upper()} cost={cost}")
+
+        self._update_span_cost(span, cost)
+
+    @debug()
+    def store_usage(self, tokens: dict = {}, span_id: Optional[str] = None) -> None:
+        """
+        ...
+        """
+        span = self._get_target_span(span_id)
+
+        logging.info(
+            f"Setting span  {span.id} {span.spankind.upper()} tokens={repr(tokens)}"
+        )
+
+        self._update_span_tokens(span, tokens)
+
     @debug()
     def dump_trace(self):
         """

{agenta-0.20.0a4 → agenta-0.20.0a6}/agenta/sdk/types.py

@@ -1,5 +1,5 @@
 import json
-from typing import Dict, List, Optional, Any
+from typing import Dict, List, Optional, Any, Union
 
 from pydantic import ConfigDict, BaseModel, HttpUrl
 
@@ -18,7 +18,7 @@ class LLMTokenUsage(BaseModel):
 
 class BaseResponse(BaseModel):
     version: Optional[str] = "2.0"
-    data: Optional[Dict[str, Any]]
+    data: Optional[Union[str, Dict[str, Any]]]
     trace: Optional[Dict[str, Any]]
 
 

{agenta-0.20.0a4 → agenta-0.20.0a6}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "agenta"
-version = "0.20.0a4"
+version = "0.20.0a6"
 description = "The SDK for agenta is an open-source LLMOps platform."
 readme = "README.md"
 authors = ["Mahmoud Mabrouk <mahmoud@agenta.ai>"]