aiqtoolkit 1.1.0a20250429__py3-none-any.whl

aiq/agent/base.py

@@ -0,0 +1,76 @@
+# 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
+from abc import ABC
+from abc import abstractmethod
+from enum import Enum
+
+from colorama import Fore
+from langchain_core.callbacks import AsyncCallbackHandler
+from langchain_core.language_models import BaseChatModel
+from langchain_core.tools import BaseTool
+from langgraph.graph.graph import CompiledGraph
+
+log = logging.getLogger(__name__)
+
+TOOL_NOT_FOUND_ERROR_MESSAGE = "There is no tool named {tool_name}. Tool must be one of {tools}."
+INPUT_SCHEMA_MESSAGE = ". Arguments must be provided as a valid JSON object following this format: {schema}"
+NO_INPUT_ERROR_MESSAGE = "No human input recieved to the agent, Please ask a valid question."
+
+AGENT_LOG_PREFIX = "[AGENT]"
+AGENT_RESPONSE_LOG_MESSAGE = f"\n{'-' * 30}\n" + \
+                                 AGENT_LOG_PREFIX + "\n" + \
+                                 Fore.YELLOW + \
+                                 "Agent input: %s\n" + \
+                                 Fore.CYAN + \
+                                 "Agent's thoughts: \n%s" + \
+                                 Fore.RESET + \
+                                 f"\n{'-' * 30}"
+
+TOOL_RESPONSE_LOG_MESSAGE = f"\n{'-' * 30}\n" + \
+                                 AGENT_LOG_PREFIX + "\n" + \
+                                 Fore.WHITE + \
+                                 "Calling tools: %s\n" + \
+                                 Fore.YELLOW + \
+                                 "Tool's input: %s\n" + \
+                                 Fore.CYAN + \
+                                 "Tool's response: \n%s" + \
+                                 Fore.RESET + \
+                                 f"\n{'-' * 30}"
+
+
+class AgentDecision(Enum):
+    TOOL = "tool"
+    END = "finished"
+
+
+class BaseAgent(ABC):
+
+    def __init__(self,
+                 llm: BaseChatModel,
+                 tools: list[BaseTool],
+                 callbacks: list[AsyncCallbackHandler] = None,
+                 detailed_logs: bool = False):
+        log.debug("Initializing Agent Graph")
+        self.llm = llm
+        self.tools = tools
+        self.callbacks = callbacks or []
+        self.detailed_logs = detailed_logs
+        self.graph = None
+
+    @abstractmethod
+    async def _build_graph(self, state_schema) -> CompiledGraph:
+        pass

aiq/agent/dual_node.py

@@ -0,0 +1,67 @@
+# 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
+from abc import abstractmethod
+
+from langchain_core.callbacks import AsyncCallbackHandler
+from langchain_core.language_models import BaseChatModel
+from langchain_core.tools import BaseTool
+from langgraph.graph import StateGraph
+from langgraph.graph.graph import CompiledGraph
+from pydantic import BaseModel
+
+from .base import AgentDecision
+from .base import BaseAgent
+
+log = logging.getLogger(__name__)
+
+
+class DualNodeAgent(BaseAgent):
+
+    def __init__(self,
+                 llm: BaseChatModel,
+                 tools: list[BaseTool],
+                 callbacks: list[AsyncCallbackHandler] = None,
+                 detailed_logs: bool = False):
+        super().__init__(llm=llm, tools=tools, callbacks=callbacks, detailed_logs=detailed_logs)
+
+    @abstractmethod
+    async def agent_node(self, state: BaseModel) -> BaseModel:
+        pass
+
+    @abstractmethod
+    async def tool_node(self, state: BaseModel) -> BaseModel:
+        pass
+
+    @abstractmethod
+    async def conditional_edge(self, state: BaseModel) -> str:
+        pass
+
+    async def _build_graph(self, state_schema) -> CompiledGraph:
+        log.debug("Building and compiling the Agent Graph")
+
+        graph = StateGraph(state_schema)
+        graph.add_node("agent", self.agent_node)
+        graph.add_node("tool", self.tool_node)
+        graph.add_edge("tool", "agent")
+
+        conditional_edge_possible_outputs = {AgentDecision.TOOL: "tool", AgentDecision.END: "__end__"}
+        graph.add_conditional_edges("agent", self.conditional_edge, conditional_edge_possible_outputs)
+
+        graph.set_entry_point("agent")
+        self.graph = graph.compile()
+
+        return self.graph

aiq/agent/react_agent/agent.py

@@ -0,0 +1,322 @@
+# 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 json
+# pylint: disable=R0917
+import logging
+from json import JSONDecodeError
+
+from langchain_core.agents import AgentAction
+from langchain_core.agents import AgentFinish
+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
+from langchain_core.runnables.config import RunnableConfig
+from langchain_core.tools import BaseTool
+from pydantic import BaseModel
+from pydantic import Field
+
+from aiq.agent.base import AGENT_LOG_PREFIX
+from aiq.agent.base import AGENT_RESPONSE_LOG_MESSAGE
+from aiq.agent.base import INPUT_SCHEMA_MESSAGE
+from aiq.agent.base import NO_INPUT_ERROR_MESSAGE
+from aiq.agent.base import TOOL_NOT_FOUND_ERROR_MESSAGE
+from aiq.agent.base import TOOL_RESPONSE_LOG_MESSAGE
+from aiq.agent.base import AgentDecision
+from aiq.agent.dual_node import DualNodeAgent
+from aiq.agent.react_agent.output_parser import ReActOutputParser
+from aiq.agent.react_agent.output_parser import ReActOutputParserException
+
+logger = logging.getLogger(__name__)
+
+
+class ReActGraphState(BaseModel):
+    """State schema for the ReAct Agent Graph"""
+    messages: list[BaseMessage] = Field(default_factory=list)  # input and output of the ReAct Agent
+    agent_scratchpad: list[AgentAction] = Field(default_factory=list)  # agent thoughts / intermediate steps
+    tool_responses: list[BaseMessage] = Field(default_factory=list)  # the responses from any tool calls
+
+
+class ReActAgentGraph(DualNodeAgent):
+    """Configurable LangGraph ReAct Agent. A ReAct Agent performs reasoning inbetween tool calls, and utilizes the tool
+    names and descriptions to select the optimal tool.  Supports retrying on output parsing errors.  Argument
+    "detailed_logs" toggles logging of inputs, outputs, and intermediate steps."""
+
+    def __init__(self,
+                 llm: BaseChatModel,
+                 prompt: ChatPromptTemplate,
+                 tools: list[BaseTool],
+                 use_tool_schema: bool = True,
+                 callbacks: list[AsyncCallbackHandler] = None,
+                 detailed_logs: bool = False,
+                 retry_parsing_errors: bool = True,
+                 max_retries: int = 1):
+        super().__init__(llm=llm, tools=tools, callbacks=callbacks, detailed_logs=detailed_logs)
+        self.retry_parsing_errors = retry_parsing_errors
+        self.max_tries = (max_retries + 1) if retry_parsing_errors else 1
+        logger.debug(
+            "%s Filling the prompt variables 'tools' and 'tool_names', using the tools provided in the config.",
+            AGENT_LOG_PREFIX)
+        tool_names = ",".join([tool.name for tool in tools[:-1]]) + ',' + tools[-1].name  # prevent trailing ","
+        if not use_tool_schema:
+            tool_names_and_descriptions = "\n".join(
+                [f"{tool.name}: {tool.description}"
+                 for tool in tools[:-1]]) + "\n" + f"{tools[-1].name}: {tools[-1].description}"  # prevent trailing "\n"
+        else:
+            logger.debug("%s Adding the tools' input schema to the tools' description", AGENT_LOG_PREFIX)
+            tool_names_and_descriptions = "\n".join([
+                f"{tool.name}: {tool.description}. {INPUT_SCHEMA_MESSAGE.format(schema=tool.input_schema.model_fields)}"
+                for tool in tools[:-1]
+            ]) + "\n" + (f"{tools[-1].name}: {tools[-1].description}. "
+                         f"{INPUT_SCHEMA_MESSAGE.format(schema=tools[-1].input_schema.model_fields)}")
+        prompt = prompt.partial(tools=tool_names_and_descriptions, tool_names=tool_names)
+        # construct the ReAct Agent
+        llm = llm.bind(stop=["Observation:"])
+        self.agent = prompt | llm
+        self.tools_dict = {tool.name: tool for tool in tools}
+        logger.debug("%s Initialized ReAct Agent Graph", AGENT_LOG_PREFIX)
+
+    def _get_tool(self, tool_name):
+        try:
+            return self.tools_dict.get(tool_name)
+        except Exception as ex:
+            logger.exception("%s Unable to find tool with the name %s\n%s",
+                             AGENT_LOG_PREFIX,
+                             tool_name,
+                             ex,
+                             exc_info=True)
+            raise ex
+
+    async def agent_node(self, state: ReActGraphState):
+        try:
+            logger.debug("%s Starting the ReAct Agent Node", AGENT_LOG_PREFIX)
+            # keeping a working state allows us to resolve parsing errors without polluting the agent scratchpad
+            # the agent "forgets" about the parsing error after solving it - prevents hallucinations in next cycles
+            working_state = []
+            for attempt in range(1, self.max_tries + 1):
+                # the first time we are invoking the ReAct Agent, it won't have any intermediate steps / agent thoughts
+                if len(state.agent_scratchpad) == 0 and len(working_state) == 0:
+                    # the user input comes from the "messages" state channel
+                    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
+                    if state.messages[0].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 = state.messages[0].content
+                    logger.debug("%s Querying agent, attempt: %s", AGENT_LOG_PREFIX, attempt)
+                    output_message = ""
+                    async for event in self.agent.astream({"question": question},
+                                                          config=RunnableConfig(callbacks=self.callbacks)):
+                        output_message += event.content
+                    output_message = AIMessage(content=output_message)
+                    if self.detailed_logs:
+                        logger.info(AGENT_RESPONSE_LOG_MESSAGE, question, output_message.content)
+                else:
+                    # ReAct Agents require agentic cycles
+                    # in an agentic cycle, preserve the agent's thoughts from the previous cycles,
+                    # and give the agent the response from the tool it called
+                    agent_scratchpad = []
+                    for index, intermediate_step in enumerate(state.agent_scratchpad):
+                        agent_thoughts = AIMessage(content=intermediate_step.log)
+                        agent_scratchpad.append(agent_thoughts)
+                        tool_response = HumanMessage(content=state.tool_responses[index].content)
+                        agent_scratchpad.append(tool_response)
+                    agent_scratchpad += working_state
+                    question = state.messages[0].content
+                    logger.debug("%s Querying agent, attempt: %s", AGENT_LOG_PREFIX, attempt)
+                    output_message = ""
+                    async for event in self.agent.astream({
+                            "question": question, "agent_scratchpad": agent_scratchpad
+                    },
+                                                          config=RunnableConfig(callbacks=self.callbacks)):
+                        output_message += event.content
+                    output_message = AIMessage(content=output_message)
+                    if self.detailed_logs:
+                        logger.info(AGENT_RESPONSE_LOG_MESSAGE, question, output_message.content)
+                        logger.debug("%s The agent's scratchpad (with tool result) was:\n%s",
+                                     AGENT_LOG_PREFIX,
+                                     agent_scratchpad)
+                try:
+                    # check if the agent has the final answer yet
+                    logger.debug("%s Successfully obtained agent response. Parsing agent's response", AGENT_LOG_PREFIX)
+                    agent_output = await ReActOutputParser().aparse(output_message.content)
+                    logger.debug("%s Successfully parsed agent's response", AGENT_LOG_PREFIX)
+                    if attempt > 1:
+                        logger.debug("%s Successfully parsed agent response after %s attempts",
+                                     AGENT_LOG_PREFIX,
+                                     attempt)
+                    if isinstance(agent_output, AgentFinish):
+                        final_answer = agent_output.return_values.get('output', output_message.content)
+                        logger.debug("%s The agent has finished, and has the final answer", AGENT_LOG_PREFIX)
+                        # this is where we handle the final output of the Agent, we can clean-up/format/postprocess here
+                        # the final answer goes in the "messages" state channel
+                        state.messages += [AIMessage(content=final_answer)]
+                    else:
+                        # the agent wants to call a tool, ensure the thoughts are preserved for the next agentic cycle
+                        agent_output.log = output_message.content
+                        logger.debug("%s The agent wants to call a tool: %s", AGENT_LOG_PREFIX, agent_output.tool)
+                        state.agent_scratchpad += [agent_output]
+                    return state
+                except ReActOutputParserException as ex:
+                    # the agent output did not meet the expected ReAct output format. This can happen for a few reasons:
+                    # the agent mentioned a tool, but already has the final answer, this can happen with Llama models
+                    #   - the ReAct Agent already has the answer, and is reflecting on how it obtained the answer
+                    # the agent might have also missed Action or Action Input in its output
+                    logger.warning("%s Error parsing agent output\nObservation:%s\nAgent Output:\n%s",
+                                   AGENT_LOG_PREFIX,
+                                   ex.observation,
+                                   output_message.content)
+                    if attempt == self.max_tries:
+                        logger.exception(
+                            "%s Failed to parse agent output after %d attempts, consider enabling or "
+                            "increasing max_retries",
+                            AGENT_LOG_PREFIX,
+                            attempt,
+                            exc_info=True)
+                        # the final answer goes in the "messages" state channel
+                        output_message.content = ex.observation + '\n' + output_message.content
+                        state.messages += [output_message]
+                        return state
+                    # retry parsing errors, if configured
+                    logger.info("%s Retrying ReAct Agent, including output parsing Observation", AGENT_LOG_PREFIX)
+                    working_state.append(output_message)
+                    working_state.append(HumanMessage(content=ex.observation))
+        except Exception as ex:
+            logger.exception("%s Failed to call agent_node: %s", AGENT_LOG_PREFIX, ex, exc_info=True)
+            raise ex
+
+    async def conditional_edge(self, state: ReActGraphState):
+        try:
+            logger.debug("%s Starting the ReAct Conditional Edge", AGENT_LOG_PREFIX)
+            if len(state.messages) > 1:
+                # the ReAct Agent has finished executing, the last agent output was AgentFinish
+                logger.debug("%s Final answer:\n%s", AGENT_LOG_PREFIX, state.messages[-1].content)
+                return AgentDecision.END
+            # else the agent wants to call a tool
+            agent_output = state.agent_scratchpad[-1]
+            logger.debug("%s The agent wants to call: %s with input: %s",
+                         AGENT_LOG_PREFIX,
+                         agent_output.tool,
+                         agent_output.tool_input)
+            return AgentDecision.TOOL
+        except Exception as ex:
+            logger.exception("Failed to determine whether agent is calling a tool: %s", ex, exc_info=True)
+            logger.warning("%s Ending graph traversal", AGENT_LOG_PREFIX)
+            return AgentDecision.END
+
+    async def tool_node(self, state: ReActGraphState):
+        try:
+            logger.debug("%s Starting the Tool Call Node", AGENT_LOG_PREFIX)
+            if len(state.agent_scratchpad) == 0:
+                raise RuntimeError('No tool input received in state: "agent_scratchpad"')
+            agent_thoughts = state.agent_scratchpad[-1]
+            # the agent can run any installed tool, simply install the tool and add it to the config file
+            requested_tool = self._get_tool(agent_thoughts.tool)
+            if not requested_tool:
+                configured_tool_names = list(self.tools_dict.keys())
+                logger.warning(
+                    "%s ReAct Agent wants to call tool %s. In the ReAct Agent's configuration within the config file,"
+                    "there is no tool with that name: %s",
+                    AGENT_LOG_PREFIX,
+                    agent_thoughts.tool,
+                    configured_tool_names)
+                tool_response = ToolMessage(name='agent_error',
+                                            tool_call_id='agent_error',
+                                            content=TOOL_NOT_FOUND_ERROR_MESSAGE.format(tool_name=agent_thoughts.tool,
+                                                                                        tools=configured_tool_names))
+                state.tool_responses += [tool_response]
+                return state
+
+            logger.debug("%s Calling tool %s with input: %s",
+                         AGENT_LOG_PREFIX,
+                         requested_tool.name,
+                         agent_thoughts.tool_input)
+
+            # Run the tool. Try to use structured input, if possible.
+            try:
+                tool_input_str = agent_thoughts.tool_input.strip().replace("'", '"')
+                tool_input_dict = json.loads(tool_input_str) if tool_input_str != 'None' else tool_input_str
+                logger.debug("%s Successfully parsed structured tool input from Action Input", AGENT_LOG_PREFIX)
+                tool_response = await requested_tool.ainvoke(tool_input_dict,
+                                                             config=RunnableConfig(callbacks=self.callbacks))
+                if self.detailed_logs:
+                    # The tool response can be very large, so we log only the first 1000 characters
+                    tool_response_str = str(tool_response)
+                    tool_response_str = tool_response_str[:1000] + "..." if len(
+                        tool_response_str) > 1000 else tool_response_str
+                    tool_response_log_message = TOOL_RESPONSE_LOG_MESSAGE % (
+                        requested_tool.name, tool_input_str, tool_response_str)
+                    logger.info(tool_response_log_message)
+            except JSONDecodeError as ex:
+                logger.warning(
+                    "%s Unable to parse structured tool input from Action Input. Using Action Input as is."
+                    "\nParsing error: %s",
+                    AGENT_LOG_PREFIX,
+                    ex,
+                    exc_info=True)
+                tool_input_str = agent_thoughts.tool_input
+                tool_response = await requested_tool.ainvoke(tool_input_str,
+                                                             config=RunnableConfig(callbacks=self.callbacks))
+
+            # some tools, such as Wikipedia, will return an empty response when no search results are found
+            if tool_response is None or tool_response == "":
+                tool_response = "The tool provided an empty response.\n"
+            # put the tool response in the graph state
+            tool_response = ToolMessage(name=agent_thoughts.tool,
+                                        tool_call_id=agent_thoughts.tool,
+                                        content=tool_response)
+            logger.debug("%s Called tool %s with input: %s\nThe tool returned: %s",
+                         AGENT_LOG_PREFIX,
+                         requested_tool.name,
+                         agent_thoughts.tool_input,
+                         tool_response.content)
+            state.tool_responses += [tool_response]
+            return state
+        except Exception as ex:
+            logger.exception("%s Failed to call tool_node: %s", AGENT_LOG_PREFIX, ex, exc_info=ex)
+            raise ex
+
+    async def build_graph(self):
+        try:
+            await super()._build_graph(state_schema=ReActGraphState)
+            logger.debug("%s ReAct Graph built and compiled successfully", AGENT_LOG_PREFIX)
+            return self.graph
+        except Exception as ex:
+            logger.exception("%s Failed to build ReAct Graph: %s", AGENT_LOG_PREFIX, ex, exc_info=ex)
+            raise ex
+
+    @staticmethod
+    def validate_system_prompt(system_prompt: str) -> bool:
+        errors = []
+        if not system_prompt:
+            errors.append("The system prompt cannot be empty.")
+        required_prompt_variables = {
+            "{tools}": "The system prompt must contain {tools} so the agent knows about configured tools.",
+            "{tool_names}": "The system prompt must contain {tool_names} so the agent knows tool names."
+        }
+        for variable_name, error_message in required_prompt_variables.items():
+            if variable_name not in system_prompt:
+                errors.append(error_message)
+        if errors:
+            error_text = "\n".join(errors)
+            logger.exception("%s %s", AGENT_LOG_PREFIX, error_text)
+            raise ValueError(error_text)
+        return True

aiq/agent/react_agent/output_parser.py

@@ -0,0 +1,104 @@
+# 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 re
+
+from langchain.agents.agent import AgentOutputParser
+from langchain_core.agents import AgentAction
+from langchain_core.agents import AgentFinish
+from langchain_core.exceptions import LangChainException
+
+from .prompt import SYSTEM_PROMPT
+
+FINAL_ANSWER_ACTION = "Final Answer:"
+MISSING_ACTION_AFTER_THOUGHT_ERROR_MESSAGE = "Invalid Format: Missing 'Action:' after 'Thought:'"
+MISSING_ACTION_INPUT_AFTER_ACTION_ERROR_MESSAGE = "Invalid Format: Missing 'Action Input:' after 'Action:'"
+FINAL_ANSWER_AND_PARSABLE_ACTION_ERROR_MESSAGE = ("Parsing LLM output produced both a final answer and a parse-able "
+                                                  "action:")
+
+
+class ReActOutputParserException(ValueError, LangChainException):
+
+    def __init__(self,
+                 observation=None,
+                 missing_action=False,
+                 missing_action_input=False,
+                 final_answer_and_action=False):
+        self.observation = observation
+        self.missing_action = missing_action
+        self.missing_action_input = missing_action_input
+        self.final_answer_and_action = final_answer_and_action
+
+
+class ReActOutputParser(AgentOutputParser):
+    """Parses ReAct-style LLM calls that have a single tool input.
+
+    Expects output to be in one of two formats.
+
+    If the output signals that an action should be taken,
+    should be in the below format. This will result in an AgentAction
+    being returned.
+
+    ```
+    Thought: agent thought here
+    Action: search
+    Action Input: what is the temperature in SF?
+    Observation: Waiting for the tool response...
+    ```
+
+    If the output signals that a final answer should be given,
+    should be in the below format. This will result in an AgentFinish
+    being returned.
+
+    ```
+    Thought: agent thought here
+    Final Answer: The temperature is 100 degrees
+    ```
+
+    """
+
+    def get_format_instructions(self) -> str:
+        return SYSTEM_PROMPT
+
+    def parse(self, text: str) -> AgentAction | AgentFinish:
+        includes_answer = FINAL_ANSWER_ACTION in text
+        regex = r"Action\s*\d*\s*:[\s]*(.*?)\s*Action\s*\d*\s*Input\s*\d*\s*:[\s]*(.*?)(?=\s*[\n|\s]\s*Observation\b|$)"
+        action_match = re.search(regex, text, re.DOTALL)
+        if action_match:
+            if includes_answer:
+                raise ReActOutputParserException(
+                    final_answer_and_action=True,
+                    observation=f"{FINAL_ANSWER_AND_PARSABLE_ACTION_ERROR_MESSAGE}: {text}")
+            action = action_match.group(1).strip()
+            action_input = action_match.group(2)
+            tool_input = action_input.strip(" ")
+            tool_input = tool_input.strip('"')
+
+            return AgentAction(action, tool_input, text)
+
+        if includes_answer:
+            return AgentFinish({"output": text.split(FINAL_ANSWER_ACTION)[-1].strip()}, text)
+
+        if not re.search(r"Action\s*\d*\s*:[\s]*(.*?)", text, re.DOTALL):
+            raise ReActOutputParserException(observation=MISSING_ACTION_AFTER_THOUGHT_ERROR_MESSAGE,
+                                             missing_action=True)
+        if not re.search(r"[\s]*Action\s*\d*\s*Input\s*\d*\s*:[\s]*(.*)", text, re.DOTALL):
+            raise ReActOutputParserException(observation=MISSING_ACTION_INPUT_AFTER_ACTION_ERROR_MESSAGE,
+                                             missing_action_input=True)
+        raise ReActOutputParserException(f"Could not parse LLM output: `{text}`")
+
+    @property
+    def _type(self) -> str:
+        return "react-input"

aiq/agent/react_agent/prompt.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.
+
+# flake8: noqa
+from langchain_core.prompts.chat import ChatPromptTemplate
+from langchain_core.prompts.chat import MessagesPlaceholder
+
+SYSTEM_PROMPT = """
+Answer the following questions as best you can. You may ask the human to use the following tools:
+
+{tools}
+
+You may respond in one of two formats.
+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")  
+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.)
+Use the following format once you have the final answer:
+
+Thought: I now know the final answer
+Final Answer: the final answer to the original input question
+"""
+USER_PROMPT = """
+Question: {question}
+"""
+
+# This is the prompt - (ReAct Agent prompt)
+react_agent_prompt = ChatPromptTemplate([("system", SYSTEM_PROMPT), ("user", USER_PROMPT),
+                                         MessagesPlaceholder(variable_name='agent_scratchpad', optional=True)])