nvidia-nat 1.2.0rc5__py3-none-any.whl

aiq/builder/eval_builder.py

@@ -0,0 +1,161 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import dataclasses
+import logging
+from contextlib import asynccontextmanager
+from pathlib import Path
+
+from aiq.builder.builder import EvalBuilder
+from aiq.builder.evaluator import EvaluatorInfo
+from aiq.builder.framework_enum import LLMFrameworkEnum
+from aiq.builder.workflow_builder import WorkflowBuilder
+from aiq.cli.type_registry import TypeRegistry
+from aiq.data_models.config import AIQConfig
+from aiq.data_models.config import GeneralConfig
+from aiq.data_models.evaluate import EvalGeneralConfig
+from aiq.data_models.evaluator import EvaluatorBaseConfig
+from aiq.data_models.function import EmptyFunctionConfig
+from aiq.utils.type_utils import override
+
+logger = logging.getLogger(__name__)
+
+
+@dataclasses.dataclass
+class ConfiguredEvaluator:
+    config: EvaluatorBaseConfig
+    instance: EvaluatorInfo
+
+
+class WorkflowEvalBuilder(WorkflowBuilder, EvalBuilder):
+
+    def __init__(self,
+                 general_config: GeneralConfig | None = None,
+                 eval_general_config: EvalGeneralConfig | None = None,
+                 registry: TypeRegistry | None = None):
+        super().__init__(general_config=general_config, registry=registry)
+        self.eval_general_config = eval_general_config
+        self._evaluators: dict[str, ConfiguredEvaluator] = {}
+
+    @override
+    async def add_evaluator(self, name: str, config: EvaluatorBaseConfig):
+        if name in self._evaluators:
+            raise ValueError(f"Evaluator `{name}` already exists in the list of evaluators")
+
+        try:
+            evaluator_info = self._registry.get_evaluator(type(config))
+            info_obj = await self._get_exit_stack().enter_async_context(evaluator_info.build_fn(config, self))
+
+            # Store the evaluator
+            self._evaluators[name] = ConfiguredEvaluator(config=config, instance=info_obj)
+        except Exception as e:
+            logger.error("Error %s adding evaluator `%s` with config `%s`", e, name, config, exc_info=True)
+            raise
+
+    @override
+    def get_evaluator(self, evaluator_name: str) -> EvaluatorInfo:
+
+        if (evaluator_name not in self._evaluators):
+            raise ValueError(f"Evaluator `{evaluator_name}` not found")
+
+        return self._evaluators[evaluator_name].instance
+
+    @override
+    def get_evaluator_config(self, evaluator_name: str) -> EvaluatorBaseConfig:
+
+        if evaluator_name not in self._evaluators:
+            raise ValueError(f"Evaluator `{evaluator_name}` not found")
+
+        # Return the tool configuration object
+        return self._evaluators[evaluator_name].config
+
+    @override
+    def get_max_concurrency(self) -> int:
+        return self.eval_general_config.max_concurrency
+
+    @override
+    def get_output_dir(self) -> Path:
+        return self.eval_general_config.output_dir
+
+    @override
+    def get_all_tools(self, wrapper_type: LLMFrameworkEnum | str):
+        tools = []
+        tool_wrapper_reg = self._registry.get_tool_wrapper(llm_framework=wrapper_type)
+        for fn_name in self._functions:
+            fn = self.get_function(fn_name)
+            try:
+                tools.append(tool_wrapper_reg.build_fn(fn_name, fn, self))
+            except Exception:
+                logger.exception("Error fetching tool `%s`", fn_name, exc_info=True)
+
+        return tools
+
+    def _log_build_failure_evaluator(self,
+                                     failing_evaluator_name: str,
+                                     completed_evaluators: list[str],
+                                     remaining_evaluators: list[str],
+                                     original_error: Exception) -> None:
+        """
+        Log comprehensive evaluator build failure information.
+
+        Args:
+            failing_evaluator_name (str): The name of the evaluator that failed to build
+            completed_evaluators (list[str]): List of evaluator names that were successfully built
+            remaining_evaluators (list[str]): List of evaluator names still to be built
+            original_error (Exception): The original exception that caused the failure
+        """
+        # Convert evaluator names to (name, type) tuples for consistent logging
+        completed_components = [(name, "evaluator") for name in completed_evaluators]
+        remaining_components = [(name, "evaluator") for name in remaining_evaluators]
+
+        # Use the inherited common logging method from WorkflowBuilder
+        self._log_build_failure(failing_evaluator_name,
+                                "evaluator",
+                                completed_components,
+                                remaining_components,
+                                original_error)
+
+    async def populate_builder(self, config: AIQConfig):
+        # Skip setting workflow if workflow config is EmptyFunctionConfig
+        skip_workflow = isinstance(config.workflow, EmptyFunctionConfig)
+
+        await super().populate_builder(config, skip_workflow)
+
+        # Initialize progress tracking for evaluators
+        completed_evaluators = []
+        remaining_evaluators = list(config.eval.evaluators.keys())
+
+        # Instantiate the evaluators with enhanced error logging
+        for name, evaluator_config in config.eval.evaluators.items():
+            try:
+                # Remove from remaining as we start building
+                remaining_evaluators.remove(name)
+
+                await self.add_evaluator(name, evaluator_config)
+
+                # Add to completed after successful build
+                completed_evaluators.append(name)
+
+            except Exception as e:
+                self._log_build_failure_evaluator(name, completed_evaluators, remaining_evaluators, e)
+                raise
+
+    @classmethod
+    @asynccontextmanager
+    async def from_config(cls, config: AIQConfig):
+
+        async with cls(config.general, config.eval.general, registry=None) as builder:
+            await builder.populate_builder(config)
+            yield builder

aiq/builder/evaluator.py

@@ -0,0 +1,29 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from collections.abc import Callable
+
+from aiq.data_models.evaluator import EvaluatorBaseConfig
+from aiq.eval.evaluator.evaluator_model import EvalInput
+from aiq.eval.evaluator.evaluator_model import EvalOutput
+
+
+class EvaluatorInfo:
+
+    def __init__(self, *, config: EvaluatorBaseConfig, evaluate_fn: Callable[[EvalInput], EvalOutput],
+                 description: str):
+        self.config = config
+        self.evaluate_fn = evaluate_fn
+        self.description = description

aiq/builder/framework_enum.py

@@ -0,0 +1,24 @@
+# SPDX-FileCopyrightText: Copyright (c) 2024-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from enum import Enum
+
+
+class LLMFrameworkEnum(str, Enum):
+    LANGCHAIN = "langchain"
+    LLAMA_INDEX = "llama_index"
+    CREWAI = "crewai"
+    SEMANTIC_KERNEL = "semantic_kernel"
+    AGNO = "agno"

aiq/builder/front_end.py

@@ -0,0 +1,73 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import typing
+from abc import ABC
+from abc import abstractmethod
+
+from aiq.data_models.front_end import FrontEndConfigT
+
+if (typing.TYPE_CHECKING):
+    from aiq.data_models.config import AIQConfig
+
+
+class FrontEndBase(typing.Generic[FrontEndConfigT], ABC):
+
+    def __init__(self, full_config: "AIQConfig"):
+        """
+        Initializes the FrontEndBase object with the specified AIQ Toolkit configuration.
+
+        Parameters
+        ----------
+        full_config : AIQConfig
+            The configuration object to use for the front end.
+        """
+
+        super().__init__()
+
+        self._full_config: "AIQConfig" = full_config
+        self._front_end_config: FrontEndConfigT = typing.cast(FrontEndConfigT, full_config.general.front_end)
+
+    @property
+    def front_end_config(self) -> FrontEndConfigT:
+        """
+        Returns the front end configuration object extracted from the AIQ Toolkit configuration.
+
+        Returns
+        -------
+        FrontEndConfigT
+            The front end configuration object.
+        """
+        return self._front_end_config
+
+    @property
+    def full_config(self) -> "AIQConfig":
+        """
+        Returns the full AIQ Toolkit configuration object.
+
+        Returns
+        -------
+        AIQConfig
+            The full AIQ Toolkit configuration object.
+        """
+
+        return self._full_config
+
+    @abstractmethod
+    async def run(self):
+        """
+        Runs the specified configuration file, launching the workflow until the front end is complete.
+        """
+        pass

aiq/builder/function.py

@@ -0,0 +1,344 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+import typing
+from abc import ABC
+from abc import abstractmethod
+from collections.abc import AsyncGenerator
+from collections.abc import Awaitable
+from collections.abc import Callable
+
+from pydantic import BaseModel
+
+from aiq.builder.context import AIQContext
+from aiq.builder.function_base import FunctionBase
+from aiq.builder.function_base import InputT
+from aiq.builder.function_base import SingleOutputT
+from aiq.builder.function_base import StreamingOutputT
+from aiq.builder.function_info import FunctionInfo
+from aiq.data_models.function import FunctionBaseConfig
+
+_InvokeFnT = Callable[[InputT], Awaitable[SingleOutputT]]
+_StreamFnT = Callable[[InputT], AsyncGenerator[StreamingOutputT]]
+
+_T = typing.TypeVar("_T")
+
+logger = logging.getLogger(__name__)
+
+
+class Function(FunctionBase[InputT, StreamingOutputT, SingleOutputT], ABC):
+
+    def __init__(self,
+                 *,
+                 config: FunctionBaseConfig,
+                 description: str | None,
+                 input_schema: type[BaseModel] | None = None,
+                 streaming_output_schema: type[BaseModel] | type[None] | None = None,
+                 single_output_schema: type[BaseModel] | type[None] | None = None,
+                 converters: list[Callable[[typing.Any], typing.Any]] | None = None,
+                 instance_name: str | None = None):
+
+        super().__init__(input_schema=input_schema,
+                         streaming_output_schema=streaming_output_schema,
+                         single_output_schema=single_output_schema,
+                         converters=converters)
+
+        self.config = config
+        self.description = description
+        self.instance_name = instance_name or config.type
+        self._context = AIQContext.get()
+
+    def convert(self, value: typing.Any, to_type: type[_T]) -> _T:
+        """
+        Converts the given value to the specified type using the function's converter.
+
+        Parameters
+        ----------
+        value : typing.Any
+            The value to convert.
+        to_type : type
+            The type to convert the value to.
+
+        Returns
+        -------
+        _T
+            The converted value.
+
+        Raises
+        ------
+        ValueError
+            If the value cannot be converted to the specified type (when `to_type` is specified).
+        """
+
+        return self._converter.convert(value, to_type=to_type)
+
+    def try_convert(self, value: typing.Any, to_type: type[_T]) -> _T | typing.Any:
+        """
+        Converts the given value to the specified type using graceful error handling.
+        If conversion fails, returns the original value and continues processing.
+
+        Parameters
+        ----------
+        value : typing.Any
+            The value to convert.
+        to_type : type
+            The type to convert the value to.
+
+        Returns
+        -------
+        _T | typing.Any
+            The converted value, or original value if conversion fails.
+        """
+        return self._converter.try_convert(value, to_type=to_type)
+
+    @abstractmethod
+    async def _ainvoke(self, value: InputT) -> SingleOutputT:
+        pass
+
+    @typing.overload
+    async def ainvoke(self, value: InputT | typing.Any) -> SingleOutputT:
+        ...
+
+    @typing.overload
+    async def ainvoke(self, value: InputT | typing.Any, to_type: type[_T]) -> _T:
+        ...
+
+    @typing.final
+    async def ainvoke(self, value: InputT | typing.Any, to_type: type | None = None):
+        """
+        Runs the function with the given input and returns a single output from the function. This is the
+        main entry point for running a function.
+
+        Parameters
+        ----------
+        value : InputT | typing.Any
+            The input to the function.
+        to_type : type | None, optional
+            The type to convert the output to using the function's converter. When not specified, the
+            output will match `single_output_type`.
+
+        Returns
+        -------
+        typing.Any
+            The output of the function optionally converted to the specified type.
+
+        Raises
+        ------
+        ValueError
+            If the output of the function cannot be converted to the specified type.
+        """
+
+        with self._context.push_active_function(self.instance_name,
+                                                input_data=value) as manager:  # Set the current invocation context
+            try:
+                converted_input: InputT = self._convert_input(value)
+
+                result = await self._ainvoke(converted_input)
+
+                if to_type is not None and not isinstance(result, to_type):
+                    result = self.convert(result, to_type)
+
+                manager.set_output(result)
+
+                return result
+            except Exception as e:
+                logger.error("Error with ainvoke in function with input: %s.", value, exc_info=True)
+                raise e
+
+    @typing.final
+    async def acall_invoke(self, *args, **kwargs):
+        """
+        A wrapper around `ainvoke` that allows for calling the function with arbitrary arguments and keyword arguments.
+        This is useful in scenarios where the function might be called by an LLM or other system which gives varying
+        inputs to the function. The function will attempt to convert the args and kwargs to the input schema of the
+        function.
+
+        Returns
+        -------
+        SingleOutputT
+            The output of the function.
+        """
+
+        if (len(args) == 1 and not kwargs):
+            # If only one argument is passed, assume it is the input just like ainvoke
+            return await self.ainvoke(value=args[0])
+
+        if (not args and kwargs):
+            # If only kwargs are passed, assume we are calling a function with named arguments in a dict
+            # This will rely on the processing in ainvoke to convert from dict to the correct input type
+            return await self.ainvoke(value=kwargs)
+
+        # Possibly have both args and kwargs, final attempt is to use the input schema object constructor.
+        try:
+            input_obj = self.input_schema(*args, **kwargs)
+
+            return await self.ainvoke(value=input_obj)
+        except Exception as e:
+            logger.error(
+                "Error in acall_invoke() converting input to function schema. Both args and kwargs were "
+                "supplied which could not be converted to the input schema. args: %s\nkwargs: %s\nschema: %s",
+                args,
+                kwargs,
+                self.input_schema)
+            raise e
+
+    @abstractmethod
+    async def _astream(self, value: InputT) -> AsyncGenerator[StreamingOutputT]:
+        yield  # type: ignore
+
+    @typing.overload
+    async def astream(self, value: InputT | typing.Any) -> AsyncGenerator[SingleOutputT]:
+        ...
+
+    @typing.overload
+    async def astream(self, value: InputT | typing.Any, to_type: type[_T]) -> AsyncGenerator[_T]:
+        ...
+
+    @typing.final
+    async def astream(self, value: InputT | typing.Any, to_type: type | None = None):
+        """
+        Runs the function with the given input and returns a stream of outputs from the function. This is the main entry
+        point for running a function with streaming output.
+
+        Parameters
+        ----------
+        value : InputT | typing.Any
+            The input to the function.
+        to_type : type | None, optional
+            The type to convert the output to using the function's converter. When not specified, the
+            output will match `streaming_output_type`.
+
+        Yields
+        ------
+        typing.Any
+            The output of the function optionally converted to the specified type.
+
+        Raises
+        ------
+        ValueError
+            If the output of the function cannot be converted to the specified type (when `to_type` is specified).
+        """
+
+        with self._context.push_active_function(self.instance_name, input_data=value) as manager:
+            try:
+                converted_input: InputT = self._convert_input(value)
+
+                # Collect streaming outputs to capture the final result
+                final_output: list[typing.Any] = []
+
+                async for data in self._astream(converted_input):
+                    if to_type is not None and not isinstance(data, to_type):
+                        converted_data = self.convert(data, to_type=to_type)
+                        final_output.append(converted_data)
+                        yield converted_data
+                    else:
+                        final_output.append(data)
+                        yield data
+
+                # Set the final output for intermediate step tracking
+                manager.set_output(final_output)
+
+            except Exception as e:
+                logger.error("Error with astream in function with input: %s.", value, exc_info=True)
+                raise e
+
+    @typing.final
+    async def acall_stream(self, *args, **kwargs):
+        """
+        A wrapper around `astream` that allows for calling the function with arbitrary arguments and keyword arguments.
+        This is useful in scenarios where the function might be called by an LLM or other system which gives varying
+        inputs to the function. The function will attempt to convert the args and kwargs to the input schema of the
+        function.
+
+        Yields
+        ------
+        StreamingOutputT
+            The output of the function.
+        """
+
+        if (len(args) == 1 and not kwargs):
+            # If only one argument is passed, assume it is the input just like ainvoke
+            async for x in self.astream(value=args[0]):
+                yield x
+
+        elif (not args and kwargs):
+            # If only kwargs are passed, assume we are calling a function with named arguments in a dict
+            # This will rely on the processing in ainvoke to convert from dict to the correct input type
+            async for x in self.astream(value=kwargs):
+                yield x
+
+        # Possibly have both args and kwargs, final attempt is to use the input schema object constructor.
+        else:
+            try:
+                input_obj = self.input_schema(*args, **kwargs)
+
+                async for x in self.astream(value=input_obj):
+                    yield x
+            except Exception as e:
+                logger.error(
+                    "Error in acall_stream() converting input to function schema. Both args and kwargs were "
+                    "supplied which could not be converted to the input schema. args: %s\nkwargs: %s\nschema: %s",
+                    args,
+                    kwargs,
+                    self.input_schema)
+                raise e
+
+
+class LambdaFunction(Function[InputT, StreamingOutputT, SingleOutputT]):
+
+    def __init__(self, *, config: FunctionBaseConfig, info: FunctionInfo, instance_name: str | None = None):
+
+        super().__init__(config=config,
+                         description=info.description,
+                         input_schema=info.input_schema,
+                         streaming_output_schema=info.stream_output_schema,
+                         single_output_schema=info.single_output_schema,
+                         converters=info.converters,
+                         instance_name=instance_name)
+
+        self._info = info
+        self._ainvoke_fn: _InvokeFnT = info.single_fn
+        self._astream_fn: _StreamFnT = info.stream_fn
+
+    @property
+    def has_streaming_output(self) -> bool:
+        return self._astream_fn is not None
+
+    @property
+    def has_single_output(self) -> bool:
+        return self._ainvoke_fn is not None
+
+    async def _ainvoke(self, value: InputT) -> SingleOutputT:
+        return await self._ainvoke_fn(value)
+
+    async def _astream(self, value: InputT) -> AsyncGenerator[StreamingOutputT]:
+        async for x in self._astream_fn(value):
+            yield x
+
+    @staticmethod
+    def from_info(*,
+                  config: FunctionBaseConfig,
+                  info: FunctionInfo,
+                  instance_name: str | None = None) -> 'LambdaFunction[InputT, StreamingOutputT, SingleOutputT]':
+
+        input_type: type = info.input_type
+        streaming_output_type = info.stream_output_type
+        single_output_type = info.single_output_type
+
+        class FunctionImpl(LambdaFunction[input_type, streaming_output_type, single_output_type]):
+            pass
+
+        return FunctionImpl(config=config, info=info, instance_name=instance_name)