nvidia-nat 1.3.dev0__py3-none-any.whl → 1.3.0.dev2__py3-none-any.whl

aiq/__init__.py

@@ -0,0 +1,66 @@
+# 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 sys
+import importlib
+import importlib.abc
+import importlib.util
+import warnings
+
+
+class CompatFinder(importlib.abc.MetaPathFinder):
+
+    def __init__(self, alias_prefix, target_prefix):
+        self.alias_prefix = alias_prefix
+        self.target_prefix = target_prefix
+
+    def find_spec(self, fullname, path, target=None):  # pylint: disable=unused-argument
+        if fullname == self.alias_prefix or fullname.startswith(self.alias_prefix + "."):
+            # Map aiq.something -> nat.something
+            target_name = self.target_prefix + fullname[len(self.alias_prefix):]
+            spec = importlib.util.find_spec(target_name)
+            if spec is None:
+                return None
+            # Wrap the loader so it loads under the alias name
+            return importlib.util.spec_from_loader(fullname, CompatLoader(fullname, target_name))
+        return None
+
+
+class CompatLoader(importlib.abc.Loader):
+
+    def __init__(self, alias_name, target_name):
+        self.alias_name = alias_name
+        self.target_name = target_name
+
+    def create_module(self, spec):
+        # Reuse the actual module so there's only one instance
+        target_module = importlib.import_module(self.target_name)
+        sys.modules[self.alias_name] = target_module
+        return target_module
+
+    def exec_module(self, module):
+        # Nothing to execute since the target is already loaded
+        pass
+
+
+# Register the compatibility finder
+sys.meta_path.insert(0, CompatFinder("aiq", "nat"))
+
+warnings.warn(
+    "!!! The 'aiq' namespace is deprecated and will be removed in a future release. "
+    "Please use the 'nat' namespace instead.",
+    DeprecationWarning,
+    stacklevel=2,
+)

nat/agent/base.py

@@ -179,6 +179,7 @@ class BaseAgent(ABC):
                 logger.debug("%s Retrying tool call for %s in %d seconds...", AGENT_LOG_PREFIX, tool.name, sleep_time)
                 await asyncio.sleep(sleep_time)
 
+        # pylint: disable=C0209
         # All retries exhausted, return error message
         error_content = "Tool call failed after all retry attempts. Last error: %s" % str(last_exception)
         logger.error("%s %s", AGENT_LOG_PREFIX, error_content)
@@ -234,6 +235,22 @@ class BaseAgent(ABC):
             logger.warning("%s Unexpected error during JSON parsing: %s", AGENT_LOG_PREFIX, str(e))
             return {"error": f"Unexpected parsing error: {str(e)}", "original_string": json_string}
 
+    def _get_chat_history(self, messages: list[BaseMessage]) -> str:
+        """
+        Get the chat history excluding the last message.
+
+        Parameters
+        ----------
+        messages : list[BaseMessage]
+            The messages to get the chat history from
+
+        Returns
+        -------
+        str
+            The chat history excluding the last message
+        """
+        return "\n".join([f"{message.type}: {message.content}" for message in messages[:-1]])
+
     @abstractmethod
     async def _build_graph(self, state_schema: type) -> CompiledGraph:
         pass

nat/agent/react_agent/agent.py

@@ -16,6 +16,7 @@
 import json
 # pylint: disable=R0917
 import logging
+import typing
 from json import JSONDecodeError
 
 from langchain_core.agents import AgentAction
@@ -44,7 +45,9 @@ from nat.agent.react_agent.output_parser import ReActOutputParser
 from nat.agent.react_agent.output_parser import ReActOutputParserException
 from nat.agent.react_agent.prompt import SYSTEM_PROMPT
 from nat.agent.react_agent.prompt import USER_PROMPT
-from nat.agent.react_agent.register import ReActAgentWorkflowConfig
+
+if typing.TYPE_CHECKING:
+    from nat.agent.react_agent.register import ReActAgentWorkflowConfig
 
 logger = logging.getLogger(__name__)
 
@@ -124,17 +127,19 @@ class ReActAgentGraph(DualNodeAgent):
                     if len(state.messages) == 0:
                         raise RuntimeError('No input received in state: "messages"')
                     # to check is any human input passed or not, if no input passed Agent will return the state
-                    content = str(state.messages[0].content)
+                    content = str(state.messages[-1].content)
                     if content.strip() == "":
                         logger.error("%s No human input passed to the agent.", AGENT_LOG_PREFIX)
                         state.messages += [AIMessage(content=NO_INPUT_ERROR_MESSAGE)]
                         return state
                     question = content
                     logger.debug("%s Querying agent, attempt: %s", AGENT_LOG_PREFIX, attempt)
-
+                    chat_history = self._get_chat_history(state.messages)
                     output_message = await self._stream_llm(
                         self.agent,
-                        {"question": question},
+                        {
+                            "question": question, "chat_history": chat_history
+                        },
                         RunnableConfig(callbacks=self.callbacks)  # type: ignore
                     )
 
@@ -152,13 +157,15 @@ class ReActAgentGraph(DualNodeAgent):
                         tool_response = HumanMessage(content=tool_response_content)
                         agent_scratchpad.append(tool_response)
                     agent_scratchpad += working_state
-                    question = str(state.messages[0].content)
+                    chat_history = self._get_chat_history(state.messages)
+                    question = str(state.messages[-1].content)
                     logger.debug("%s Querying agent, attempt: %s", AGENT_LOG_PREFIX, attempt)
 
-                    output_message = await self._stream_llm(self.agent, {
-                        "question": question, "agent_scratchpad": agent_scratchpad
-                    },
-                                                            RunnableConfig(callbacks=self.callbacks))
+                    output_message = await self._stream_llm(
+                        self.agent, {
+                            "question": question, "agent_scratchpad": agent_scratchpad, "chat_history": chat_history
+                        },
+                        RunnableConfig(callbacks=self.callbacks))
 
                     if self.detailed_logs:
                         logger.info(AGENT_CALL_LOG_MESSAGE, question, output_message.content)
@@ -326,7 +333,7 @@ class ReActAgentGraph(DualNodeAgent):
         return True
 
 
-def create_react_agent_prompt(config: ReActAgentWorkflowConfig) -> ChatPromptTemplate:
+def create_react_agent_prompt(config: "ReActAgentWorkflowConfig") -> ChatPromptTemplate:
     """
     Create a ReAct Agent prompt from the config.
 

nat/agent/react_agent/prompt.py

@@ -26,7 +26,7 @@ Use the following format exactly to ask the human to use a tool:
 Question: the input question you must answer
 Thought: you should always think about what to do
 Action: the action to take, should be one of [{tool_names}]
-Action Input: the input to the action (if there is no required input, include "Action Input: None")  
+Action Input: the input to the action (if there is no required input, include "Action Input: None")
 Observation: wait for the human to respond with the result from the tool, do not assume the response
 
 ... (this Thought/Action/Action Input/Observation can repeat N times. If you do not need to use a tool, or after asking the human to use any tools and waiting for the human to respond, you might know the final answer.)
@@ -37,5 +37,8 @@ Final Answer: the final answer to the original input question
 """
 
 USER_PROMPT = """
+Previous conversation history:
+{chat_history}
+
 Question: {question}
 """

nat/agent/rewoo_agent/agent.py

@@ -21,6 +21,7 @@ from json import JSONDecodeError
 from langchain_core.callbacks.base import AsyncCallbackHandler
 from langchain_core.language_models import BaseChatModel
 from langchain_core.messages.ai import AIMessage
+from langchain_core.messages.base import BaseMessage
 from langchain_core.messages.human import HumanMessage
 from langchain_core.messages.tool import ToolMessage
 from langchain_core.prompts.chat import ChatPromptTemplate
@@ -43,6 +44,7 @@ logger = logging.getLogger(__name__)
 
 class ReWOOGraphState(BaseModel):
     """State schema for the ReWOO Agent Graph"""
+    messages: list[BaseMessage] = Field(default_factory=list)  # input and output of the ReWOO Agent
     task: HumanMessage = Field(default_factory=lambda: HumanMessage(content=""))  # the task provided by user
     plan: AIMessage = Field(
         default_factory=lambda: AIMessage(content=""))  # the plan generated by the planner to solve the task
@@ -183,10 +185,12 @@ class ReWOOAgentGraph(BaseAgent):
             if not task:
                 logger.error("%s No task provided to the ReWOO Agent. Please provide a valid task.", AGENT_LOG_PREFIX)
                 return {"result": NO_INPUT_ERROR_MESSAGE}
-
+            chat_history = self._get_chat_history(state.messages)
             plan = await self._stream_llm(
                 planner,
-                {"task": task},
+                {
+                    "task": task, "chat_history": chat_history
+                },
                 RunnableConfig(callbacks=self.callbacks)  # type: ignore
             )
 

nat/agent/rewoo_agent/prompt.py

@@ -87,6 +87,9 @@ Begin!
 """
 
 PLANNER_USER_PROMPT = """
+Previous conversation history:
+{chat_history}
+
 task: {task}
 """
 

nat/agent/rewoo_agent/register.py

@@ -124,8 +124,9 @@ async def rewoo_agent_workflow(config: ReWOOAgentWorkflowConfig, builder: Builde
                                                         token_counter=len,
                                                         start_on="human",
                                                         include_system=True)
-            task = HumanMessage(content=messages[0].content)
-            state = ReWOOGraphState(task=task)
+
+            task = HumanMessage(content=messages[-1].content)
+            state = ReWOOGraphState(messages=messages, task=task)
 
             # run the ReWOO Agent Graph
             state = await graph.ainvoke(state)

nat/agent/tool_calling_agent/agent.py

@@ -15,11 +15,14 @@
 
 # pylint: disable=R0917
 import logging
+import typing
 
 from langchain_core.callbacks.base import AsyncCallbackHandler
 from langchain_core.language_models import BaseChatModel
+from langchain_core.messages import SystemMessage
 from langchain_core.messages.base import BaseMessage
-from langchain_core.runnables import RunnableConfig
+from langchain_core.runnables import RunnableLambda
+from langchain_core.runnables.config import RunnableConfig
 from langchain_core.tools import BaseTool
 from langgraph.prebuilt import ToolNode
 from pydantic import BaseModel
@@ -30,6 +33,9 @@ from nat.agent.base import AGENT_LOG_PREFIX
 from nat.agent.base import AgentDecision
 from nat.agent.dual_node import DualNodeAgent
 
+if typing.TYPE_CHECKING:
+    from nat.agent.tool_calling_agent.register import ToolCallAgentWorkflowConfig
+
 logger = logging.getLogger(__name__)
 
 
@@ -43,22 +49,51 @@ class ToolCallAgentGraph(DualNodeAgent):
     A tool Calling Agent utilizes the tool input parameters to select the optimal tool.  Supports handling tool errors.
     Argument "detailed_logs" toggles logging of inputs, outputs, and intermediate steps."""
 
-    def __init__(self,
-                 llm: BaseChatModel,
-                 tools: list[BaseTool],
-                 callbacks: list[AsyncCallbackHandler] = None,
-                 detailed_logs: bool = False,
-                 handle_tool_errors: bool = True):
+    def __init__(
+        self,
+        llm: BaseChatModel,
+        tools: list[BaseTool],
+        prompt: str | None = None,
+        callbacks: list[AsyncCallbackHandler] = None,
+        detailed_logs: bool = False,
+        handle_tool_errors: bool = True,
+    ):
         super().__init__(llm=llm, tools=tools, callbacks=callbacks, detailed_logs=detailed_logs)
+        # some LLMs support tool calling
+        # these models accept the tool's input schema and decide when to use a tool based on the input's relevance
+        try:
+            # in tool calling agents, we bind the tools to the LLM, to pass the tools' input schemas at runtime
+            self.bound_llm = llm.bind_tools(tools)
+        except NotImplementedError as ex:
+            logger.error("%s Failed to bind tools: %s", AGENT_LOG_PREFIX, ex, exc_info=True)
+            raise ex
+
+        if prompt is not None:
+            system_prompt = SystemMessage(content=prompt)
+            prompt_runnable = RunnableLambda(
+                lambda state: [system_prompt] + state.get("messages", []),
+                name="SystemPrompt",
+            )
+        else:
+            prompt_runnable = RunnableLambda(
+                lambda state: state.get("messages", []),
+                name="PromptPassthrough",
+            )
+
+        self.agent = prompt_runnable | self.bound_llm
+
         self.tool_caller = ToolNode(tools, handle_tool_errors=handle_tool_errors)
         logger.debug("%s Initialized Tool Calling Agent Graph", AGENT_LOG_PREFIX)
 
     async def agent_node(self, state: ToolCallAgentGraphState):
         try:
-            logger.debug('%s Starting the Tool Calling Agent Node', AGENT_LOG_PREFIX)
+            logger.debug("%s Starting the Tool Calling Agent Node", AGENT_LOG_PREFIX)
             if len(state.messages) == 0:
                 raise RuntimeError('No input received in state: "messages"')
-            response = await self.llm.ainvoke(state.messages, config=RunnableConfig(callbacks=self.callbacks))
+            response = await self.agent.ainvoke(
+                {"messages": state.messages},
+                config=RunnableConfig(callbacks=self.callbacks),
+            )
             if self.detailed_logs:
                 agent_input = "\n".join(str(message.content) for message in state.messages)
                 logger.info(AGENT_CALL_LOG_MESSAGE, agent_input, response)
@@ -75,16 +110,18 @@ class ToolCallAgentGraph(DualNodeAgent):
             last_message = state.messages[-1]
             if last_message.tool_calls:
                 # the agent wants to call a tool
-                logger.debug('%s Agent is calling a tool', AGENT_LOG_PREFIX)
+                logger.debug("%s Agent is calling a tool", AGENT_LOG_PREFIX)
                 return AgentDecision.TOOL
             if self.detailed_logs:
                 logger.debug("%s Final answer:\n%s", AGENT_LOG_PREFIX, state.messages[-1].content)
             return AgentDecision.END
         except Exception as ex:
-            logger.exception("%s Failed to determine whether agent is calling a tool: %s",
-                             AGENT_LOG_PREFIX,
-                             ex,
-                             exc_info=True)
+            logger.exception(
+                "%s Failed to determine whether agent is calling a tool: %s",
+                AGENT_LOG_PREFIX,
+                ex,
+                exc_info=True,
+            )
             logger.warning("%s Ending graph traversal", AGENT_LOG_PREFIX)
             return AgentDecision.END
 
@@ -92,14 +129,15 @@ class ToolCallAgentGraph(DualNodeAgent):
         try:
             logger.debug("%s Starting Tool Node", AGENT_LOG_PREFIX)
             tool_calls = state.messages[-1].tool_calls
-            tools = [tool.get('name') for tool in tool_calls]
+            tools = [tool.get("name") for tool in tool_calls]
             tool_input = state.messages[-1]
-            tool_response = await self.tool_caller.ainvoke(input={"messages": [tool_input]},
-                                                           config=RunnableConfig(callbacks=self.callbacks,
-                                                                                 configurable={}))
+            tool_response = await self.tool_caller.ainvoke(
+                input={"messages": [tool_input]},
+                config=RunnableConfig(callbacks=self.callbacks, configurable={}),
+            )
             # this configurable = {} argument is needed due to a bug in LangGraph PreBuilt ToolNode ^
 
-            for response in tool_response.get('messages'):
+            for response in tool_response.get("messages"):
                 if self.detailed_logs:
                     self._log_tool_response(str(tools), str(tool_input), response.content)
                 state.messages += [response]
@@ -112,8 +150,41 @@ class ToolCallAgentGraph(DualNodeAgent):
     async def build_graph(self):
         try:
             await super()._build_graph(state_schema=ToolCallAgentGraphState)
-            logger.debug("%s Tool Calling Agent Graph built and compiled successfully", AGENT_LOG_PREFIX)
+            logger.debug(
+                "%s Tool Calling Agent Graph built and compiled successfully",
+                AGENT_LOG_PREFIX,
+            )
             return self.graph
         except Exception as ex:
-            logger.exception("%s Failed to build Tool Calling Agent Graph: %s", AGENT_LOG_PREFIX, ex, exc_info=ex)
+            logger.exception(
+                "%s Failed to build Tool Calling Agent Graph: %s",
+                AGENT_LOG_PREFIX,
+                ex,
+                exc_info=ex,
+            )
             raise ex
+
+
+def create_tool_calling_agent_prompt(config: "ToolCallAgentWorkflowConfig") -> str | None:
+    """
+    Create a Tool Calling Agent prompt from the config.
+
+    Args:
+        config (ToolCallAgentWorkflowConfig): The config to use for the prompt.
+
+    Returns:
+        ChatPromptTemplate: The Tool Calling Agent prompt.
+    """
+    # the Tool Calling Agent prompt can be customized via config option system_prompt and additional_instructions.
+
+    if config.system_prompt:
+        prompt_str = config.system_prompt
+    else:
+        prompt_str = ""
+
+    if config.additional_instructions:
+        prompt_str += f" {config.additional_instructions}"
+
+    if len(prompt_str) > 0:
+        return prompt_str
+    return None

nat/agent/tool_calling_agent/register.py

@@ -41,6 +41,9 @@ class ToolCallAgentWorkflowConfig(FunctionBaseConfig, name="tool_calling_agent")
     handle_tool_errors: bool = Field(default=True, description="Specify ability to handle tool calling errors.")
     description: str = Field(default="Tool Calling Agent Workflow", description="Description of this functions use.")
     max_iterations: int = Field(default=15, description="Number of tool calls before stoping the tool calling agent.")
+    system_prompt: str | None = Field(default=None, description="Provides the system prompt to use with the agent.")
+    additional_instructions: str | None = Field(default=None,
+                                                description="Additional instructions appended to the system prompt.")
 
 
 @register_function(config_type=ToolCallAgentWorkflowConfig, framework_wrappers=[LLMFrameworkEnum.LANGCHAIN])
@@ -49,10 +52,11 @@ async def tool_calling_agent_workflow(config: ToolCallAgentWorkflowConfig, build
     from langgraph.graph.graph import CompiledGraph
 
     from nat.agent.base import AGENT_LOG_PREFIX
+    from nat.agent.tool_calling_agent.agent import ToolCallAgentGraph
+    from nat.agent.tool_calling_agent.agent import ToolCallAgentGraphState
+    from nat.agent.tool_calling_agent.agent import create_tool_calling_agent_prompt
 
-    from .agent import ToolCallAgentGraph
-    from .agent import ToolCallAgentGraphState
-
+    prompt = create_tool_calling_agent_prompt(config)
     # we can choose an LLM for the ReAct agent in the config file
     llm = await builder.get_llm(config.llm_name, wrapper_type=LLMFrameworkEnum.LANGCHAIN)
     # the agent can run any installed tool, simply install the tool and add it to the config file
@@ -61,18 +65,10 @@ async def tool_calling_agent_workflow(config: ToolCallAgentWorkflowConfig, build
     if not tools:
         raise ValueError(f"No tools specified for Tool Calling Agent '{config.llm_name}'")
 
-    # some LLMs support tool calling
-    # these models accept the tool's input schema and decide when to use a tool based on the input's relevance
-    try:
-        # in tool calling agents, we bind the tools to the LLM, to pass the tools' input schemas at runtime
-        llm = llm.bind_tools(tools)
-    except NotImplementedError as ex:
-        logger.error("%s Failed to bind tools: %s", AGENT_LOG_PREFIX, ex, exc_info=True)
-        raise ex
-
     # construct the Tool Calling Agent Graph from the configured llm, and tools
     graph: CompiledGraph = await ToolCallAgentGraph(llm=llm,
                                                     tools=tools,
+                                                    prompt=prompt,
                                                     detailed_logs=config.verbose,
                                                     handle_tool_errors=config.handle_tool_errors).build_graph()
 

nat/cli/type_registry.py

@@ -588,8 +588,8 @@ class TypeRegistry:  # pylint: disable=too-many-public-methods
         except KeyError as err:
             raise KeyError(
                 f"An invalid Embedder config and wrapper combination was supplied. Config: `{config_type}`, "
-                "Wrapper: `{wrapper_type}`. The workflow is requesting a {wrapper_type} Embedder client but "
-                "there is no registered conversion from that Embedder provider to LLM framework: {wrapper_type}. "
+                f"Wrapper: `{wrapper_type}`. The workflow is requesting a {wrapper_type} Embedder client but "
+                f"there is no registered conversion from that Embedder provider to LLM framework: {wrapper_type}. "
                 "Please provide an Embedder configuration from one of the following providers: "
                 f"{set(self._embedder_client_provider_to_framework.keys())}") from err
 
@@ -703,8 +703,8 @@ class TypeRegistry:  # pylint: disable=too-many-public-methods
         except KeyError as err:
             raise KeyError(
                 f"An invalid Retriever config and wrapper combination was supplied. Config: `{config_type}`, "
-                "Wrapper: `{wrapper_type}`. The workflow is requesting a {wrapper_type} Retriever client but "
-                "there is no registered conversion from that Retriever provider to LLM framework: {wrapper_type}. "
+                f"Wrapper: `{wrapper_type}`. The workflow is requesting a {wrapper_type} Retriever client but "
+                f"there is no registered conversion from that Retriever provider to LLM framework: {wrapper_type}. "
                 "Please provide a Retriever configuration from one of the following providers: "
                 f"{set(self._retriever_client_provider_to_framework.keys())}") from err
 

nat/embedder/azure_openai_embedder.py

@@ -0,0 +1,46 @@
+# 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 pydantic import AliasChoices
+from pydantic import ConfigDict
+from pydantic import Field
+
+from nat.builder.builder import Builder
+from nat.builder.embedder import EmbedderProviderInfo
+from nat.cli.register_workflow import register_embedder_provider
+from nat.data_models.embedder import EmbedderBaseConfig
+from nat.data_models.retry_mixin import RetryMixin
+
+
+class AzureOpenAIEmbedderModelConfig(EmbedderBaseConfig, RetryMixin, name="azure_openai"):
+    """An Azure OpenAI embedder provider to be used with an embedder client."""
+
+    model_config = ConfigDict(protected_namespaces=(), extra="allow")
+
+    api_key: str | None = Field(default=None, description="Azure OpenAI API key to interact with hosted model.")
+    api_version: str = Field(default="2025-04-01-preview", description="Azure OpenAI API version.")
+    azure_endpoint: str | None = Field(validation_alias=AliasChoices("azure_endpoint", "base_url"),
+                                       serialization_alias="azure_endpoint",
+                                       default=None,
+                                       description="Base URL for the hosted model.")
+    azure_deployment: str = Field(validation_alias=AliasChoices("azure_deployment", "model_name", "model"),
+                                  serialization_alias="azure_deployment",
+                                  description="The Azure OpenAI hosted model/deployment name.")
+
+
+@register_embedder_provider(config_type=AzureOpenAIEmbedderModelConfig)
+async def azure_openai_embedder_model(config: AzureOpenAIEmbedderModelConfig, _builder: Builder):
+
+    yield EmbedderProviderInfo(config=config, description="An Azure OpenAI model for use with an Embedder client.")

nat/embedder/openai_embedder.py

@@ -34,10 +34,9 @@ class OpenAIEmbedderModelConfig(EmbedderBaseConfig, RetryMixin, name="openai"):
     model_name: str = Field(validation_alias=AliasChoices("model_name", "model"),
                             serialization_alias="model",
                             description="The OpenAI hosted model name.")
-    max_retries: int = Field(default=2, description="The max number of retries for the request.")
 
 
 @register_embedder_provider(config_type=OpenAIEmbedderModelConfig)
-async def openai_llm(config: OpenAIEmbedderModelConfig, builder: Builder):
+async def openai_embedder_model(config: OpenAIEmbedderModelConfig, _builder: Builder):
 
     yield EmbedderProviderInfo(config=config, description="An OpenAI model for use with an Embedder client.")

nat/embedder/register.py

@@ -18,5 +18,6 @@
 # isort:skip_file
 
 # Import any providers which need to be automatically registered here
+from . import azure_openai_embedder
 from . import nim_embedder
 from . import openai_embedder

nat/llm/azure_openai_llm.py

@@ -0,0 +1,50 @@
+# 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 pydantic import AliasChoices
+from pydantic import ConfigDict
+from pydantic import Field
+
+from nat.builder.builder import Builder
+from nat.builder.llm import LLMProviderInfo
+from nat.cli.register_workflow import register_llm_provider
+from nat.data_models.llm import LLMBaseConfig
+from nat.data_models.retry_mixin import RetryMixin
+
+
+class AzureOpenAIModelConfig(LLMBaseConfig, RetryMixin, name="azure_openai"):
+    """An Azure OpenAI LLM provider to be used with an LLM client."""
+
+    model_config = ConfigDict(protected_namespaces=(), extra="allow")
+
+    api_key: str | None = Field(default=None, description="Azure OpenAI API key to interact with hosted model.")
+    api_version: str = Field(default="2025-04-01-preview", description="Azure OpenAI API version.")
+    azure_endpoint: str | None = Field(validation_alias=AliasChoices("azure_endpoint", "base_url"),
+                                       serialization_alias="azure_endpoint",
+                                       default=None,
+                                       description="Base URL for the hosted model.")
+    azure_deployment: str = Field(validation_alias=AliasChoices("azure_deployment", "model_name", "model"),
+                                  serialization_alias="azure_deployment",
+                                  description="The Azure OpenAI hosted model/deployment name.")
+    temperature: float = Field(default=0.0, description="Sampling temperature in [0, 1].")
+    top_p: float = Field(default=1.0, description="Top-p for distribution sampling.")
+    seed: int | None = Field(default=None, description="Random seed to set for generation.")
+    max_retries: int = Field(default=10, description="The max number of retries for the request.")
+
+
+@register_llm_provider(config_type=AzureOpenAIModelConfig)
+async def azure_openai_llm(config: AzureOpenAIModelConfig, _builder: Builder):
+
+    yield LLMProviderInfo(config=config, description="An Azure OpenAI model for use with an LLM client.")

nat/llm/register.py

@@ -19,5 +19,6 @@
 
 # Import any providers which need to be automatically registered here
 from . import aws_bedrock_llm
+from . import azure_openai_llm
 from . import nim_llm
 from . import openai_llm

nat/meta/pypi.md

@@ -23,19 +23,19 @@ NeMo Agent toolkit is a flexible library designed to seamlessly integrate your e
 
 ## Key Features
 
-- [**Framework Agnostic:**](https://docs.nvidia.com/nemo/agent-toolkit/1.2.0/extend/plugins.html) Works with any agentic framework, so you can use your current technology stack without replatforming.
-- [**Reusability:**](https://docs.nvidia.com/nemo/agent-toolkit/1.2.0/extend/sharing-components.html) Every agent, tool, or workflow can be combined and repurposed, allowing developers to leverage existing work in new scenarios.
-- [**Rapid Development:**](https://docs.nvidia.com/nemo/agent-toolkit/1.2.0/tutorials/index.html) Start with a pre-built agent, tool, or workflow, and customize it to your needs.
-- [**Profiling:**](https://docs.nvidia.com/nemo/agent-toolkit/1.2.0/workflows/profiler.html) Profile entire workflows down to the tool and agent level, track input/output tokens and timings, and identify bottlenecks.
-- [**Observability:**](https://docs.nvidia.com/nemo/agent-toolkit/1.2.0/workflows/observe/observe-workflow-with-phoenix.html) Monitor and debug your workflows with any OpenTelemetry-compatible observability tool, with examples using [Phoenix](https://docs.nvidia.com/nemo/agent-toolkit/1.2.0/workflows/observe/observe-workflow-with-phoenix.html) and [W&B Weave](https://docs.nvidia.com/nemo/agent-toolkit/1.2.0/workflows/observe/observe-workflow-with-weave.html).
-- [**Evaluation System:**](https://docs.nvidia.com/nemo/agent-toolkit/1.2.0/workflows/evaluate.html) Validate and maintain accuracy of agentic workflows with built-in evaluation tools.
-- [**User Interface:**](https://docs.nvidia.com/nemo/agent-toolkit/1.2.0/quick-start/launching-ui.html) Use the NeMo Agent toolkit UI chat interface to interact with your agents, visualize output, and debug workflows.
-- [**MCP Compatibility**](https://docs.nvidia.com/nemo/agent-toolkit/1.2.0/workflows/mcp/mcp-client.html) Compatible with Model Context Protocol (MCP), allowing tools served by MCP Servers to be used as NeMo Agent toolkit functions.
+- [**Framework Agnostic:**](https://docs.nvidia.com/nemo/agent-toolkit/1.2/extend/plugins.html) Works with any agentic framework, so you can use your current technology stack without replatforming.
+- [**Reusability:**](https://docs.nvidia.com/nemo/agent-toolkit/1.2/extend/sharing-components.html) Every agent, tool, or workflow can be combined and repurposed, allowing developers to leverage existing work in new scenarios.
+- [**Rapid Development:**](https://docs.nvidia.com/nemo/agent-toolkit/1.2/tutorials/index.html) Start with a pre-built agent, tool, or workflow, and customize it to your needs.
+- [**Profiling:**](https://docs.nvidia.com/nemo/agent-toolkit/1.2/workflows/profiler.html) Profile entire workflows down to the tool and agent level, track input/output tokens and timings, and identify bottlenecks.
+- [**Observability:**](https://docs.nvidia.com/nemo/agent-toolkit/1.2/workflows/observe/observe-workflow-with-phoenix.html) Monitor and debug your workflows with any OpenTelemetry-compatible observability tool, with examples using [Phoenix](https://docs.nvidia.com/nemo/agent-toolkit/1.2/workflows/observe/observe-workflow-with-phoenix.html) and [W&B Weave](https://docs.nvidia.com/nemo/agent-toolkit/1.2/workflows/observe/observe-workflow-with-weave.html).
+- [**Evaluation System:**](https://docs.nvidia.com/nemo/agent-toolkit/1.2/workflows/evaluate.html) Validate and maintain accuracy of agentic workflows with built-in evaluation tools.
+- [**User Interface:**](https://docs.nvidia.com/nemo/agent-toolkit/1.2/quick-start/launching-ui.html) Use the NeMo Agent toolkit UI chat interface to interact with your agents, visualize output, and debug workflows.
+- [**MCP Compatibility**](https://docs.nvidia.com/nemo/agent-toolkit/1.2/workflows/mcp/mcp-client.html) Compatible with Model Context Protocol (MCP), allowing tools served by MCP Servers to be used as NeMo Agent toolkit functions.
 
 With NeMo Agent toolkit, you can move quickly, experiment freely, and ensure reliability across all your agent-driven projects.
 
 ## Links
- * [Documentation](https://docs.nvidia.com/nemo/agent-toolkit/1.2.0/index.html): Explore the full documentation for NeMo Agent toolkit.
+ * [Documentation](https://docs.nvidia.com/nemo/agent-toolkit/1.2/index.html): Explore the full documentation for NeMo Agent toolkit.
 
 ## First time user?
  If this is your first time using NeMo Agent toolkit, it is recommended to install the latest version from the [source repository](https://github.com/NVIDIA/NeMo-Agent-Toolkit?tab=readme-ov-file#quick-start) on GitHub. This package is intended for users who are familiar with NeMo Agent toolkit applications and need to add NeMo Agent toolkit as a dependency to their project.