agenta 0.19.8a0__tar.gz → 0.20.0a0__tar.gz

{agenta-0.19.8a0 → agenta-0.20.0a0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: agenta
-Version: 0.19.8a0
+Version: 0.20.0a0
 Summary: The SDK for agenta is an open-source LLMOps platform.
 Home-page: https://agenta.ai
 Keywords: LLMOps,LLM,evaluation,prompt engineering

{agenta-0.19.8a0 → agenta-0.20.0a0}/agenta/__init__.py

@@ -17,7 +17,7 @@ from .sdk.types import (
 from .sdk.tracing.logger import llm_logger as logging
 from .sdk.tracing.llm_tracing import Tracing
 from .sdk.decorators.tracing import instrument
-from .sdk.decorators.llm_entrypoint import entrypoint, app
+from .sdk.decorators.llm_entrypoint import entrypoint, app, route
 from .sdk.agenta_init import Config, AgentaSingleton, init
 from .sdk.utils.helper.openai_cost import calculate_token_usage
 from .sdk.client import Agenta

{agenta-0.19.8a0 → agenta-0.20.0a0}/agenta/client/backend/types/create_span.py

@@ -20,7 +20,10 @@ class CreateSpan(pydantic.BaseModel):
     variant_id: typing.Optional[str]
     variant_name: typing.Optional[str]
     inputs: typing.Optional[typing.Dict[str, typing.Any]]
-    outputs: typing.Optional[typing.List[str]]
+    internals: typing.Optional[typing.Dict[str, typing.Any]]
+    outputs: typing.Optional[
+        typing.Union[typing.Dict[str, typing.Any], typing.List[str]]
+    ]
     config: typing.Optional[typing.Dict[str, typing.Any]]
     environment: typing.Optional[str]
     tags: typing.Optional[typing.List[str]]

{agenta-0.19.8a0 → agenta-0.20.0a0}/agenta/client/backend/types/llm_tokens.py

@@ -12,9 +12,9 @@ except ImportError:
 
 
 class LlmTokens(pydantic.BaseModel):
-    prompt_tokens: typing.Optional[int]
-    completion_tokens: typing.Optional[int]
-    total_tokens: typing.Optional[int]
+    prompt_tokens: typing.Optional[int] = 0
+    completion_tokens: typing.Optional[int] = 0
+    total_tokens: typing.Optional[int] = 0
 
     def json(self, **kwargs: typing.Any) -> str:
         kwargs_with_defaults: typing.Any = {

{agenta-0.19.8a0 → agenta-0.20.0a0}/agenta/sdk/__init__.py

@@ -16,7 +16,7 @@ from .types import (
 
 from .tracing.llm_tracing import Tracing
 from .decorators.tracing import instrument
-from .decorators.llm_entrypoint import entrypoint, app
+from .decorators.llm_entrypoint import entrypoint, app, route
 from .agenta_init import Config, AgentaSingleton, init
 from .utils.helper.openai_cost import calculate_token_usage
 

{agenta-0.19.8a0 → agenta-0.20.0a0}/agenta/sdk/decorators/llm_entrypoint.py

@@ -3,6 +3,7 @@
 import os
 import sys
 import time
+import json
 import inspect
 import argparse
 import asyncio
@@ -15,11 +16,11 @@ from typing import Any, Callable, Dict, Optional, Tuple, List
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi import Body, FastAPI, UploadFile, HTTPException
 
-import agenta
+import agenta as ag
 from agenta.sdk.context import save_context
 from agenta.sdk.router import router as router
 from agenta.sdk.tracing.logger import llm_logger as logging
-from agenta.sdk.tracing.llm_tracing import Tracing
+from agenta.sdk.tracing.tracing_context import tracing_context, TracingContext
 from agenta.sdk.decorators.base import BaseDecorator
 from agenta.sdk.types import (
     Context,
@@ -32,10 +33,12 @@ from agenta.sdk.types import (
     TextParam,
     MessagesInput,
     FileInputURL,
-    FuncResponse,
+    BaseResponse,
     BinaryParam,
 )
 
+from pydantic import BaseModel, HttpUrl
+
 app = FastAPI()
 
 origins = [
@@ -52,16 +55,56 @@ app.add_middleware(
 
 app.include_router(router, prefix="")
 
-
 from agenta.sdk.utils.debug import debug, DEBUG, SHIFT
 
 
 logging.setLevel("DEBUG")
 
 
+class PathValidator(BaseModel):
+    url: HttpUrl
+
+
+class route(BaseDecorator):
+    # This decorator is used to expose specific stages of a workflow (embedding, retrieval, summarization, etc.)
+    # as independent endpoints. It is designed for backward compatibility with existing code that uses
+    # the @entrypoint decorator, which has certain limitations. By using @route(), we can create new
+    # routes without altering the main workflow entrypoint. This helps in modularizing the services
+    # and provides flexibility in how we expose different functionalities as APIs.
+    def __init__(self, path):
+        path = "/" + path.strip("/").strip()
+        path = "" if path == "/" else path
+
+        PathValidator(url=f"http://example.com{path}")
+
+        self.route_path = path
+
+    def __call__(self, f):
+        self.e = entrypoint(f, route_path=self.route_path)
+
+        return f
+
+
 class entrypoint(BaseDecorator):
-    """Decorator class to wrap a function for HTTP POST, terminal exposure and enable tracing.
+    """
+    Decorator class to wrap a function for HTTP POST, terminal exposure and enable tracing.
+
+    This decorator generates the following endpoints:
+
+    Playground Endpoints
+    - /generate                 with @entrypoint, @route("/"), @route(path="") # LEGACY
+    - /playground/run           with @entrypoint, @route("/"), @route(path="")
+    - /playground/run/{route}   with @route({route}), @route(path={route})
 
+    Deployed Endpoints:
+    - /generate_deployed        with @entrypoint, @route("/"), @route(path="") # LEGACY
+    - /run                      with @entrypoint, @route("/"), @route(path="")
+    - /run/{route}              with @route({route}), @route(path={route})
+
+    The rationale is:
+    - There may be multiple endpoints, based on the different routes.
+    - It's better to make it explicit that an endpoint is for the playground.
+    - Prefixing the routes with /run is more futureproof in case we add more endpoints.
 
     Example:
     ```python
@@ -73,31 +116,64 @@ class entrypoint(BaseDecorator):
     ```
     """
 
-    def __init__(self, func: Callable[..., Any]):
-        endpoint_name = "generate"
+    routes = list()
+
+    def __init__(self, func: Callable[..., Any], route_path=""):
+        DEFAULT_PATH = "generate"
+        PLAYGROUND_PATH = "/playground"
+        RUN_PATH = "/run"
+
         func_signature = inspect.signature(func)
-        config_params = agenta.config.all()
+        config_params = ag.config.all()
         ingestible_files = self.extract_ingestible_files(func_signature)
 
+        ### --- Playground  --- #
         @debug()
         @functools.wraps(func)
         async def wrapper(*args, **kwargs) -> Any:
             func_params, api_config_params = self.split_kwargs(kwargs, config_params)
             self.ingest_files(func_params, ingestible_files)
-            agenta.config.set(**api_config_params)
+            ag.config.set(**api_config_params)
 
             # Set the configuration and environment of the LLM app parent span at run-time
-            agenta.tracing.update_baggage(
+            ag.tracing.update_baggage(
                 {"config": config_params, "environment": "playground"}
             )
 
-            # Exceptions are all handled inside self.execute_function()
-            llm_result = await self.execute_function(
+            entrypoint_result = await self.execute_function(
                 func, *args, params=func_params, config_params=config_params
             )
 
-            return llm_result
+            return entrypoint_result
+
+        self.update_function_signature(
+            wrapper, func_signature, config_params, ingestible_files
+        )
 
+        #
+        if route_path == "":
+            route = f"/{DEFAULT_PATH}"
+            app.post(route, response_model=BaseResponse)(wrapper)
+            entrypoint.routes.append(
+                {
+                    "func": func.__name__,
+                    "endpoint": DEFAULT_PATH,
+                    "params": {**config_params, **func_signature.parameters},
+                }
+            )
+
+        route = f"{PLAYGROUND_PATH}{RUN_PATH}{route_path}"
+        app.post(route, response_model=BaseResponse)(wrapper)
+        entrypoint.routes.append(
+            {
+                "func": func.__name__,
+                "endpoint": route[1:].replace("/", "_"),
+                "params": {**config_params, **func_signature.parameters},
+            }
+        )
+        ### ---------------------------- #
+
+        ### --- Deployed / Published --- #
         @debug()
         @functools.wraps(func)
         async def wrapper_deployed(*args, **kwargs) -> Any:
@@ -106,44 +182,51 @@ class entrypoint(BaseDecorator):
             }
 
             if "environment" in kwargs and kwargs["environment"] is not None:
-                agenta.config.pull(environment_name=kwargs["environment"])
+                ag.config.pull(environment_name=kwargs["environment"])
             elif "config" in kwargs and kwargs["config"] is not None:
-                agenta.config.pull(config_name=kwargs["config"])
+                ag.config.pull(config_name=kwargs["config"])
             else:
-                agenta.config.pull(config_name="default")
+                ag.config.pull(config_name="default")
 
             # Set the configuration and environment of the LLM app parent span at run-time
-            agenta.tracing.update_baggage(
+            ag.tracing.update_baggage(
                 {"config": config_params, "environment": kwargs["environment"]}
             )
 
-            llm_result = await self.execute_function(
+            entrypoint_result = await self.execute_function(
                 func, *args, params=func_params, config_params=config_params
             )
 
-            return llm_result
-
-        self.update_function_signature(
-            wrapper, func_signature, config_params, ingestible_files
-        )
-        route = f"/{endpoint_name}"
-        app.post(route, response_model=FuncResponse)(wrapper)
+            return entrypoint_result
 
         self.update_deployed_function_signature(
             wrapper_deployed,
             func_signature,
             ingestible_files,
         )
-        route_deployed = f"/{endpoint_name}_deployed"
-        app.post(route_deployed, response_model=FuncResponse)(wrapper_deployed)
-        self.override_schema(
-            openapi_schema=app.openapi(),
-            func_name=func.__name__,
-            endpoint=endpoint_name,
-            params={**config_params, **func_signature.parameters},
-        )
 
-        if self.is_main_script(func):
+        if route_path == "/":
+            route_deployed = f"/{DEFAULT_PATH}_deployed"
+            app.post(route_deployed, response_model=BaseResponse)(wrapper_deployed)
+
+        route_deployed = f"{RUN_PATH}{route_path}"
+        app.post(route_deployed, response_model=BaseResponse)(wrapper_deployed)
+        ### ---------------------------- #
+
+        ### --- Update OpenAPI --- #
+        app.openapi_schema = None  # Forces FastAPI to re-generate the schema
+        openapi_schema = app.openapi()
+
+        for route in entrypoint.routes:
+            self.override_schema(
+                openapi_schema=openapi_schema,
+                func=route["func"],
+                endpoint=route["endpoint"],
+                params=route["params"],
+            )
+        ### ---------------------- #
+
+        if self.is_main_script(func) and route_path == "":
             self.handle_terminal_run(
                 func,
                 func_signature.parameters,  # type: ignore
@@ -198,51 +281,55 @@ class entrypoint(BaseDecorator):
             For synchronous functions, it calls them directly, while for asynchronous functions,
             it awaits their execution.
             """
+            data = None
+            trace = None
+
+            token = None
+            if tracing_context.get() is None:
+                token = tracing_context.set(TracingContext())
+
             is_coroutine_function = inspect.iscoroutinefunction(func)
-            start_time = time.perf_counter()
 
             if is_coroutine_function:
                 result = await func(*args, **func_params["params"])
             else:
                 result = func(*args, **func_params["params"])
 
-            end_time = time.perf_counter()
-            latency = end_time - start_time
+            if token is not None:
+                trace = ag.tracing.dump_trace()
+                tracing_context.reset(token)
 
             if isinstance(result, Context):
                 save_context(result)
+
             if isinstance(result, Dict):
-                return FuncResponse(**result, latency=round(latency, 4))
-            if isinstance(result, str):
-                return FuncResponse(
-                    message=result, usage=None, cost=None, latency=round(latency, 4)
-                )
-            if isinstance(result, int) or isinstance(result, float):
-                return FuncResponse(
-                    message=str(result),
-                    usage=None,
-                    cost=None,
-                    latency=round(latency, 4),
-                )
-            if result is None:
-                return FuncResponse(
-                    message="Function executed successfully, but did return None. \n Are you sure you did not forget to return a value?",
-                    usage=None,
-                    cost=None,
-                    latency=round(latency, 4),
+                data = result
+            elif isinstance(result, str):
+                data = {"message": result}
+            elif isinstance(result, int) or isinstance(result, float):
+                data = {"message": str(result)}
+
+            if data is None:
+                warning = (
+                    "Function executed successfully, but did return None. \n Are you sure you did not forget to return a value?",
                 )
+
+                data = {"message": warning}
+
+            return BaseResponse(data=data, trace=trace)
+
         except Exception as e:
             self.handle_exception(e)
-        return FuncResponse(message="Unexpected error occurred when calling the @entrypoint decorated function", latency=0)  # type: ignore
 
     def handle_exception(self, e: Exception):
-        """Handle exceptions."""
+        status_code = e.status_code if hasattr(e, "status_code") else 500
+        message = str(e)
+        stacktrace = traceback.format_exception(e, value=e, tb=e.__traceback__)  # type: ignore
+        detail = {"message": message, "stacktrace": stacktrace}
 
-        status_code: int = e.status_code if hasattr(e, "status_code") else 500
-        traceback_str = traceback.format_exception(e, value=e, tb=e.__traceback__)  # type: ignore
         raise HTTPException(
             status_code=status_code,
-            detail={"error": str(e), "traceback": "".join(traceback_str)},
+            detail=detail,
         )
 
     def update_wrapper_signature(
@@ -418,26 +505,29 @@ class entrypoint(BaseDecorator):
                 file_path=args_func_params[name],
             )
 
-        agenta.config.set(**args_config_params)
+        ag.config.set(**args_config_params)
 
         # Set the configuration and environment of the LLM app parent span at run-time
-        agenta.tracing.update_baggage(
-            {"config": agenta.config.all(), "environment": "bash"}
-        )
+        ag.tracing.update_baggage({"config": ag.config.all(), "environment": "bash"})
 
         loop = asyncio.get_event_loop()
+
         result = loop.run_until_complete(
             self.execute_function(
                 func,
                 **{"params": args_func_params, "config_params": args_config_params},
             )
         )
-        print(
-            f"\n========== Result ==========\n\nMessage: {result.message}\nCost: {result.cost}\nToken Usage: {result.usage}"
-        )
+
+        print(f"\n========== Result ==========\n")
+
+        print("-> data")
+        print(json.dumps(result.data, indent=2))
+        print("-> trace")
+        print(json.dumps(result.trace, indent=2))
 
     def override_schema(
-        self, openapi_schema: dict, func_name: str, endpoint: str, params: dict
+        self, openapi_schema: dict, func: str, endpoint: str, params: dict
     ):
         """
         Overrides the default openai schema generated by fastapi with additional information about:
@@ -452,7 +542,7 @@ class entrypoint(BaseDecorator):
 
         Args:
             openapi_schema (dict): The openapi schema generated by fastapi
-            func_name (str): The name of the function to override
+            func (str): The name of the function to override
             endpoint (str): The name of the endpoint to override
             params (dict(param_name, param_val)): The dictionary of the parameters for the function
         """
@@ -484,8 +574,29 @@ class entrypoint(BaseDecorator):
                         # value = {'temperature': { "type": "number", "title": "Temperature", "x-parameter": "float" }}
                     return value
 
+        def get_type_from_param(param_val):
+            param_type = "string"
+            annotation = param_val.annotation
+
+            if annotation == int:
+                param_type = "integer"
+            elif annotation == float:
+                param_type = "number"
+            elif annotation == dict:
+                param_type = "object"
+            elif annotation == bool:
+                param_type = "boolean"
+            elif annotation == list:
+                param_type = "list"
+            elif annotation == str:
+                param_type = "string"
+            else:
+                print("ERROR, unhandled annotation:", annotation)
+
+            return param_type
+
         schema_to_override = openapi_schema["components"]["schemas"][
-            f"Body_{func_name}_{endpoint}_post"
+            f"Body_{func}_{endpoint}_post"
         ]["properties"]
         for param_name, param_val in params.items():
             if isinstance(param_val, GroupedMultipleChoiceParam):
@@ -501,7 +612,7 @@ class entrypoint(BaseDecorator):
                 subschema["choices"] = param_val.choices  # type: ignore
                 subschema["default"] = param_val.default  # type: ignore
 
-            if isinstance(param_val, MultipleChoiceParam):
+            elif isinstance(param_val, MultipleChoiceParam):
                 subschema = find_in_schema(
                     param_val.__schema_type_properties__(),
                     schema_to_override,
@@ -520,7 +631,7 @@ class entrypoint(BaseDecorator):
                     default if default in param_choices else choices[0]
                 )
 
-            if isinstance(param_val, FloatParam):
+            elif isinstance(param_val, FloatParam):
                 subschema = find_in_schema(
                     param_val.__schema_type_properties__(),
                     schema_to_override,
@@ -531,7 +642,7 @@ class entrypoint(BaseDecorator):
                 subschema["maximum"] = param_val.maxval  # type: ignore
                 subschema["default"] = param_val
 
-            if isinstance(param_val, IntParam):
+            elif isinstance(param_val, IntParam):
                 subschema = find_in_schema(
                     param_val.__schema_type_properties__(),
                     schema_to_override,
@@ -542,7 +653,7 @@ class entrypoint(BaseDecorator):
                 subschema["maximum"] = param_val.maxval  # type: ignore
                 subschema["default"] = param_val
 
-            if (
+            elif (
                 isinstance(param_val, inspect.Parameter)
                 and param_val.annotation is DictInput
             ):
@@ -554,7 +665,7 @@ class entrypoint(BaseDecorator):
                 )
                 subschema["default"] = param_val.default["default_keys"]
 
-            if isinstance(param_val, TextParam):
+            elif isinstance(param_val, TextParam):
                 subschema = find_in_schema(
                     param_val.__schema_type_properties__(),
                     schema_to_override,
@@ -563,7 +674,7 @@ class entrypoint(BaseDecorator):
                 )
                 subschema["default"] = param_val
 
-            if (
+            elif (
                 isinstance(param_val, inspect.Parameter)
                 and param_val.annotation is MessagesInput
             ):
@@ -575,7 +686,7 @@ class entrypoint(BaseDecorator):
                 )
                 subschema["default"] = param_val.default
 
-            if (
+            elif (
                 isinstance(param_val, inspect.Parameter)
                 and param_val.annotation is FileInputURL
             ):
@@ -587,7 +698,7 @@ class entrypoint(BaseDecorator):
                 )
                 subschema["default"] = "https://example.com"
 
-            if isinstance(param_val, BinaryParam):
+            elif isinstance(param_val, BinaryParam):
                 subschema = find_in_schema(
                     param_val.__schema_type_properties__(),
                     schema_to_override,
@@ -595,3 +706,11 @@ class entrypoint(BaseDecorator):
                     "bool",
                 )
                 subschema["default"] = param_val.default  # type: ignore
+            else:
+                subschema = {
+                    "title": str(param_name).capitalize(),
+                    "type": get_type_from_param(param_val),
+                }
+                if param_val.default != inspect._empty:
+                    subschema["default"] = param_val.default  # type: ignore
+                schema_to_override[param_name] = subschema

agenta-0.20.0a0/agenta/sdk/decorators/tracing.py

@@ -0,0 +1,91 @@
+# Stdlib Imports
+import inspect
+import traceback
+from functools import wraps
+from typing import Any, Callable, Optional
+
+# Own Imports
+import agenta as ag
+from agenta.sdk.decorators.base import BaseDecorator
+from agenta.sdk.tracing.logger import llm_logger as logging
+from agenta.sdk.utils.debug import debug, DEBUG, SHIFT
+
+
+logging.setLevel("DEBUG")
+
+
+class instrument(BaseDecorator):
+    """Decorator class for monitoring llm apps functions.
+
+    Args:
+        BaseDecorator (object): base decorator class
+
+    Example:
+    ```python
+        import agenta as ag
+
+        prompt_config = {"system_prompt": ..., "temperature": 0.5, "max_tokens": ...}
+
+        @ag.instrument(spankind="llm")
+        async def litellm_openai_call(prompt:str) -> str:
+            return "do something"
+
+        @ag.instrument(config=prompt_config) # spankind for parent span defaults to workflow
+        async def generate(prompt: str):
+            return ...
+    ```
+    """
+
+    def __init__(
+        self,
+        config: Optional[dict] = None,
+        spankind: str = "workflow",
+    ) -> None:
+        self.config = config
+        self.spankind = spankind
+
+    def __call__(self, func: Callable[..., Any]):
+        is_coroutine_function = inspect.iscoroutinefunction(func)
+
+        def get_inputs(*args, **kwargs):
+            func_args = inspect.getfullargspec(func).args
+            input_dict = {name: value for name, value in zip(func_args, args)}
+            input_dict.update(kwargs)
+
+            return input_dict
+
+        @wraps(func)
+        async def async_wrapper(*args, **kwargs):
+            async def wrapped_func(*args, **kwargs):
+                with ag.tracing.Context(
+                    name=func.__name__,
+                    input=get_inputs(*args, **kwargs),
+                    spankind=self.spankind,
+                    config=self.config,
+                ):
+                    result = await func(*args, **kwargs)
+
+                    ag.tracing.store_outputs(result)
+
+                    return result
+
+            return await wrapped_func(*args, **kwargs)
+
+        @wraps(func)
+        def sync_wrapper(*args, **kwargs):
+            def wrapped_func(*args, **kwargs):
+                with ag.tracing.Context(
+                    name=func.__name__,
+                    input=get_inputs(*args, **kwargs),
+                    spankind=self.spankind,
+                    config=self.config,
+                ):
+                    result = func(*args, **kwargs)
+
+                    ag.tracing.store_outputs(result)
+
+                    return result
+
+            return wrapped_func(*args, **kwargs)
+
+        return async_wrapper if is_coroutine_function else sync_wrapper