uipath-langchain 0.0.133__py3-none-any.whl → 0.1.24__py3-none-any.whl

uipath_langchain/agent/guardrails/guardrails_subgraph.py

@@ -0,0 +1,247 @@
+from typing import Any, Callable, Sequence
+
+from langgraph.constants import END, START
+from langgraph.graph import StateGraph
+from uipath.platform.guardrails import (
+    BaseGuardrail,
+    BuiltInValidatorGuardrail,
+    GuardrailScope,
+)
+
+from uipath_langchain.agent.guardrails.types import ExecutionStage
+
+from .actions.base_action import GuardrailAction, GuardrailActionNode
+from .guardrail_nodes import (
+    create_agent_guardrail_node,
+    create_llm_guardrail_node,
+    create_tool_guardrail_node,
+)
+from .types import AgentGuardrailsGraphState
+
+_VALIDATOR_ALLOWED_STAGES = {
+    "prompt_injection": {ExecutionStage.PRE_EXECUTION},
+    "pii_detection": {ExecutionStage.PRE_EXECUTION, ExecutionStage.POST_EXECUTION},
+}
+
+
+def _filter_guardrails_by_stage(
+    guardrails: Sequence[tuple[BaseGuardrail, GuardrailAction]] | None,
+    stage: ExecutionStage,
+) -> list[tuple[BaseGuardrail, GuardrailAction]]:
+    """Filter guardrails that apply to a specific execution stage."""
+    filtered_guardrails = []
+    for guardrail, action in guardrails or []:
+        # Internal knowledge: Check against configured allowed stages
+        if (
+            isinstance(guardrail, BuiltInValidatorGuardrail)
+            and guardrail.validator_type in _VALIDATOR_ALLOWED_STAGES
+            and stage not in _VALIDATOR_ALLOWED_STAGES[guardrail.validator_type]
+        ):
+            continue
+        filtered_guardrails.append((guardrail, action))
+    return filtered_guardrails
+
+
+def _create_guardrails_subgraph(
+    main_inner_node: tuple[str, Any],
+    guardrails: Sequence[tuple[BaseGuardrail, GuardrailAction]] | None,
+    scope: GuardrailScope,
+    execution_stages: Sequence[ExecutionStage],
+    node_factory: Callable[
+        [
+            BaseGuardrail,
+            ExecutionStage,
+            str,  # success node name
+            str,  # fail node name
+        ],
+        GuardrailActionNode,
+    ] = create_llm_guardrail_node,
+):
+    """Build a subgraph that enforces guardrails around an inner node.
+
+    The constructed graph conditionally includes pre- and/or post-execution guardrail
+    chains based on ``execution_stages``:
+    - If ``ExecutionStage.PRE_EXECUTION`` is included, the graph links
+      START -> first pre-guardrail node -> ... -> inner.
+      Otherwise, it directly links START -> inner.
+    - If ``ExecutionStage.POST_EXECUTION`` is included, the graph links
+      inner -> first post-guardrail node -> ... -> END.
+      Otherwise, it directly links inner -> END.
+
+    No static edges are added between guardrail nodes; each evaluation node routes
+    dynamically to its configured success/failure targets. Failure nodes are added
+    but not chained; they are expected to route via Command to the provided next node.
+    """
+    inner_name, inner_node = main_inner_node
+
+    subgraph = StateGraph(AgentGuardrailsGraphState)
+
+    subgraph.add_node(inner_name, inner_node)
+
+    # Add pre execution guardrail nodes
+    if ExecutionStage.PRE_EXECUTION in execution_stages:
+        pre_guardrails = _filter_guardrails_by_stage(
+            guardrails, ExecutionStage.PRE_EXECUTION
+        )
+        first_pre_exec_guardrail_node = _build_guardrail_node_chain(
+            subgraph,
+            pre_guardrails,
+            scope,
+            ExecutionStage.PRE_EXECUTION,
+            node_factory,
+            inner_name,
+        )
+        subgraph.add_edge(START, first_pre_exec_guardrail_node)
+    else:
+        subgraph.add_edge(START, inner_name)
+
+    # Add post execution guardrail nodes
+    if ExecutionStage.POST_EXECUTION in execution_stages:
+        post_guardrails = _filter_guardrails_by_stage(
+            guardrails, ExecutionStage.POST_EXECUTION
+        )
+        first_post_exec_guardrail_node = _build_guardrail_node_chain(
+            subgraph,
+            post_guardrails,
+            scope,
+            ExecutionStage.POST_EXECUTION,
+            node_factory,
+            END,
+        )
+        subgraph.add_edge(inner_name, first_post_exec_guardrail_node)
+    else:
+        subgraph.add_edge(inner_name, END)
+
+    return subgraph.compile()
+
+
+def _build_guardrail_node_chain(
+    subgraph: StateGraph[AgentGuardrailsGraphState],
+    guardrails: Sequence[tuple[BaseGuardrail, GuardrailAction]] | None,
+    scope: GuardrailScope,
+    execution_stage: ExecutionStage,
+    node_factory: Callable[
+        [
+            BaseGuardrail,
+            ExecutionStage,
+            str,  # success node name
+            str,  # fail node name
+        ],
+        GuardrailActionNode,
+    ],
+    next_node: str,
+) -> str:
+    """Recursively build a chain of guardrail nodes in reverse order.
+
+    This function processes guardrails from last to first, creating a chain where:
+    - Each guardrail node evaluates the guardrail condition
+    - On success, it routes to the next guardrail node (or the final next_node)
+    - On failure, it routes to a failure node that either throws an error or continues to next_node
+
+    Args:
+        subgraph: The StateGraph to add nodes and edges to.
+        guardrails: Sequence of (guardrail, action) tuples to process. Processed in reverse.
+        scope: The scope of the guardrails (LLM, AGENT, or TOOL).
+        execution_stage: Whether this is "PreExecution" or "PostExecution" guardrails.
+        node_factory: Factory function to create guardrail evaluation nodes.
+        next_node: The node name to route to after all guardrails pass.
+
+    Returns:
+        The name of the first guardrail node in the chain (or next_node if no guardrails).
+    """
+    # Base case: no guardrails to process, return the next node directly
+    if not guardrails:
+        return next_node
+
+    guardrail, action = guardrails[-1]
+    remaining_guardrails = guardrails[:-1]
+
+    fail_node_name, fail_node = action.action_node(
+        guardrail=guardrail, scope=scope, execution_stage=execution_stage
+    )
+
+    # Create the guardrail evaluation node.
+    guardrail_node_name, guardrail_node = node_factory(
+        guardrail, execution_stage, next_node, fail_node_name
+    )
+
+    # Add both nodes to the subgraph
+    subgraph.add_node(guardrail_node_name, guardrail_node)
+    subgraph.add_node(fail_node_name, fail_node)
+
+    # Failure path route to the next node
+    subgraph.add_edge(fail_node_name, next_node)
+
+    previous_node_name = _build_guardrail_node_chain(
+        subgraph,
+        remaining_guardrails,
+        scope,
+        execution_stage,
+        node_factory,
+        guardrail_node_name,
+    )
+
+    return previous_node_name
+
+
+def create_llm_guardrails_subgraph(
+    llm_node: tuple[str, Any],
+    guardrails: Sequence[tuple[BaseGuardrail, GuardrailAction]] | None,
+):
+    applicable_guardrails = [
+        (guardrail, _)
+        for (guardrail, _) in (guardrails or [])
+        if GuardrailScope.LLM in guardrail.selector.scopes
+    ]
+    return _create_guardrails_subgraph(
+        main_inner_node=llm_node,
+        guardrails=applicable_guardrails,
+        scope=GuardrailScope.LLM,
+        execution_stages=[ExecutionStage.PRE_EXECUTION, ExecutionStage.POST_EXECUTION],
+        node_factory=create_llm_guardrail_node,
+    )
+
+
+def create_agent_guardrails_subgraph(
+    agent_node: tuple[str, Any],
+    guardrails: Sequence[tuple[BaseGuardrail, GuardrailAction]] | None,
+    execution_stage: ExecutionStage,
+):
+    """Create a subgraph for AGENT-scoped guardrails that applies checks at the specified stage.
+
+    This is intended for wrapping nodes like INIT or TERMINATE, where guardrails should run
+    either before (pre-execution) or after (post-execution) the node logic.
+    """
+    applicable_guardrails = [
+        (guardrail, _)
+        for (guardrail, _) in (guardrails or [])
+        if GuardrailScope.AGENT in guardrail.selector.scopes
+    ]
+    return _create_guardrails_subgraph(
+        main_inner_node=agent_node,
+        guardrails=applicable_guardrails,
+        scope=GuardrailScope.AGENT,
+        execution_stages=[execution_stage],
+        node_factory=create_agent_guardrail_node,
+    )
+
+
+def create_tool_guardrails_subgraph(
+    tool_node: tuple[str, Any],
+    guardrails: Sequence[tuple[BaseGuardrail, GuardrailAction]] | None,
+):
+    tool_name, _ = tool_node
+    applicable_guardrails = [
+        (guardrail, action)
+        for (guardrail, action) in (guardrails or [])
+        if GuardrailScope.TOOL in guardrail.selector.scopes
+        and guardrail.selector.match_names is not None
+        and tool_name in guardrail.selector.match_names
+    ]
+    return _create_guardrails_subgraph(
+        main_inner_node=tool_node,
+        guardrails=applicable_guardrails,
+        scope=GuardrailScope.TOOL,
+        execution_stages=[ExecutionStage.PRE_EXECUTION, ExecutionStage.POST_EXECUTION],
+        node_factory=create_tool_guardrail_node,
+    )

uipath_langchain/agent/guardrails/types.py

@@ -0,0 +1,20 @@
+from enum import Enum
+from typing import Annotated, Optional
+
+from langchain_core.messages import AnyMessage
+from langgraph.graph.message import add_messages
+from pydantic import BaseModel
+
+
+class AgentGuardrailsGraphState(BaseModel):
+    """Agent Guardrails Graph state for guardrail subgraph."""
+
+    messages: Annotated[list[AnyMessage], add_messages] = []
+    guardrail_validation_result: Optional[str] = None
+
+
+class ExecutionStage(str, Enum):
+    """Execution stage enumeration."""
+
+    PRE_EXECUTION = "preExecution"
+    POST_EXECUTION = "postExecution"

uipath_langchain/agent/react/__init__.py

@@ -0,0 +1,14 @@
+"""UiPath ReAct Agent implementation"""
+
+from .agent import create_agent
+from .types import AgentGraphConfig, AgentGraphNode, AgentGraphState
+from .utils import resolve_input_model, resolve_output_model
+
+__all__ = [
+    "create_agent",
+    "resolve_output_model",
+    "resolve_input_model",
+    "AgentGraphNode",
+    "AgentGraphState",
+    "AgentGraphConfig",
+]

uipath_langchain/agent/react/agent.py

@@ -0,0 +1,113 @@
+import os
+from typing import Callable, Sequence, Type, TypeVar, cast
+
+from langchain_core.language_models import BaseChatModel
+from langchain_core.messages import HumanMessage, SystemMessage
+from langchain_core.tools import BaseTool
+from langgraph.constants import END, START
+from langgraph.graph import StateGraph
+from pydantic import BaseModel
+from uipath.platform.guardrails import BaseGuardrail
+
+from ..guardrails import create_llm_guardrails_subgraph
+from ..guardrails.actions import GuardrailAction
+from ..tools import create_tool_node
+from .init_node import (
+    create_init_node,
+)
+from .llm_node import (
+    create_llm_node,
+)
+from .router import (
+    route_agent,
+)
+from .terminate_node import (
+    create_terminate_node,
+)
+from .tools import create_flow_control_tools
+from .types import AgentGraphConfig, AgentGraphNode, AgentGraphState
+
+InputT = TypeVar("InputT", bound=BaseModel)
+OutputT = TypeVar("OutputT", bound=BaseModel)
+
+
+def create_state_with_input(input_schema: Type[InputT]):
+    InnerAgentGraphState = type(
+        "InnerAgentGraphState",
+        (AgentGraphState, input_schema),
+        {},
+    )
+
+    cast(type[BaseModel], InnerAgentGraphState).model_rebuild()
+    return InnerAgentGraphState
+
+
+def create_agent(
+    model: BaseChatModel,
+    tools: Sequence[BaseTool],
+    messages: Sequence[SystemMessage | HumanMessage]
+    | Callable[[InputT], Sequence[SystemMessage | HumanMessage]],
+    *,
+    input_schema: Type[InputT] | None = None,
+    output_schema: Type[OutputT] | None = None,
+    config: AgentGraphConfig | None = None,
+    guardrails: Sequence[tuple[BaseGuardrail, GuardrailAction]] | None = None,
+) -> StateGraph[AgentGraphState, None, InputT, OutputT]:
+    """Build agent graph with INIT -> AGENT(subgraph) <-> TOOLS loop, terminated by control flow tools.
+
+    The AGENT node is a subgraph that runs:
+    - before-agent guardrail middlewares
+    - the LLM tool-executing node
+    - after-agent guardrail middlewares
+
+    Control flow tools (end_execution, raise_error) are auto-injected alongside regular tools.
+    """
+    if config is None:
+        config = AgentGraphConfig()
+
+    os.environ["LANGCHAIN_RECURSION_LIMIT"] = str(config.recursion_limit)
+
+    agent_tools = list(tools)
+    flow_control_tools: list[BaseTool] = create_flow_control_tools(output_schema)
+    llm_tools: list[BaseTool] = [*agent_tools, *flow_control_tools]
+
+    init_node = create_init_node(messages)
+    tool_nodes = create_tool_node(agent_tools)
+    terminate_node = create_terminate_node(output_schema)
+
+    InnerAgentGraphState = create_state_with_input(
+        input_schema if input_schema is not None else BaseModel
+    )
+
+    builder: StateGraph[AgentGraphState, None, InputT, OutputT] = StateGraph(
+        InnerAgentGraphState, input_schema=input_schema, output_schema=output_schema
+    )
+    builder.add_node(AgentGraphNode.INIT, init_node)
+
+    for tool_name, tool_node in tool_nodes.items():
+        builder.add_node(tool_name, tool_node)
+
+    builder.add_node(AgentGraphNode.TERMINATE, terminate_node)
+
+    builder.add_edge(START, AgentGraphNode.INIT)
+
+    llm_node = create_llm_node(model, llm_tools)
+    llm_with_guardrails_subgraph = create_llm_guardrails_subgraph(
+        (AgentGraphNode.LLM, llm_node), guardrails
+    )
+    builder.add_node(AgentGraphNode.AGENT, llm_with_guardrails_subgraph)
+    builder.add_edge(AgentGraphNode.INIT, AgentGraphNode.AGENT)
+
+    tool_node_names = list(tool_nodes.keys())
+    builder.add_conditional_edges(
+        AgentGraphNode.AGENT,
+        route_agent,
+        [AgentGraphNode.AGENT, *tool_node_names, AgentGraphNode.TERMINATE],
+    )
+
+    for tool_name in tool_node_names:
+        builder.add_edge(tool_name, AgentGraphNode.AGENT)
+
+    builder.add_edge(AgentGraphNode.TERMINATE, END)
+
+    return builder

uipath_langchain/agent/react/constants.py

@@ -0,0 +1,2 @@
+# Agent routing configuration
+MAX_SUCCESSIVE_COMPLETIONS = 1

uipath_langchain/agent/react/init_node.py

@@ -0,0 +1,20 @@
+"""State initialization node for the ReAct Agent graph."""
+
+from typing import Any, Callable, Sequence
+
+from langchain_core.messages import HumanMessage, SystemMessage
+
+
+def create_init_node(
+    messages: Sequence[SystemMessage | HumanMessage]
+    | Callable[[Any], Sequence[SystemMessage | HumanMessage]],
+):
+    def graph_state_init(state: Any):
+        if callable(messages):
+            resolved_messages = messages(state)
+        else:
+            resolved_messages = messages
+
+        return {"messages": list(resolved_messages)}
+
+    return graph_state_init

uipath_langchain/agent/react/llm_node.py

@@ -0,0 +1,43 @@
+"""LLM node implementation for LangGraph."""
+
+from typing import Sequence
+
+from langchain_core.language_models import BaseChatModel
+from langchain_core.messages import AIMessage, AnyMessage
+from langchain_core.tools import BaseTool
+
+from .constants import MAX_SUCCESSIVE_COMPLETIONS
+from .types import AgentGraphState
+from .utils import count_successive_completions
+
+
+def create_llm_node(
+    model: BaseChatModel,
+    tools: Sequence[BaseTool] | None = None,
+):
+    """Invoke LLM with tools and dynamically control tool_choice based on successive completions.
+
+    When successive completions reach the limit, tool_choice is set to "required" to force
+    the LLM to use a tool and prevent infinite reasoning loops.
+    """
+    bindable_tools = list(tools) if tools else []
+    base_llm = model.bind_tools(bindable_tools) if bindable_tools else model
+
+    async def llm_node(state: AgentGraphState):
+        messages: list[AnyMessage] = state.messages
+
+        successive_completions = count_successive_completions(messages)
+        if successive_completions >= MAX_SUCCESSIVE_COMPLETIONS:
+            llm = base_llm.bind(tool_choice="required")
+        else:
+            llm = base_llm
+
+        response = await llm.ainvoke(messages)
+        if not isinstance(response, AIMessage):
+            raise TypeError(
+                f"LLM returned {type(response).__name__} instead of AIMessage"
+            )
+
+        return {"messages": [response]}
+
+    return llm_node

uipath_langchain/agent/react/router.py

@@ -0,0 +1,97 @@
+"""Routing functions for conditional edges in the agent graph."""
+
+from typing import Literal
+
+from langchain_core.messages import AIMessage, AnyMessage, ToolCall
+from uipath.agent.react import END_EXECUTION_TOOL, RAISE_ERROR_TOOL
+
+from ..exceptions import AgentNodeRoutingException
+from .constants import MAX_SUCCESSIVE_COMPLETIONS
+from .types import AgentGraphNode, AgentGraphState
+from .utils import count_successive_completions
+
+FLOW_CONTROL_TOOLS = [END_EXECUTION_TOOL.name, RAISE_ERROR_TOOL.name]
+
+
+def __filter_control_flow_tool_calls(
+    tool_calls: list[ToolCall],
+) -> list[ToolCall]:
+    """Remove control flow tools when multiple tool calls exist."""
+    if len(tool_calls) <= 1:
+        return tool_calls
+
+    return [tc for tc in tool_calls if tc.get("name") not in FLOW_CONTROL_TOOLS]
+
+
+def __has_control_flow_tool(tool_calls: list[ToolCall]) -> bool:
+    """Check if any tool call is of a control flow tool."""
+    return any(tc.get("name") in FLOW_CONTROL_TOOLS for tc in tool_calls)
+
+
+def __validate_last_message_is_AI(messages: list[AnyMessage]) -> AIMessage:
+    """Validate and return last message from state.
+
+    Raises:
+        AgentNodeRoutingException: If messages are empty or last message is not AIMessage
+    """
+    if not messages:
+        raise AgentNodeRoutingException(
+            "No messages in state - cannot route after agent"
+        )
+
+    last_message = messages[-1]
+    if not isinstance(last_message, AIMessage):
+        raise AgentNodeRoutingException(
+            f"Last message is not AIMessage (type: {type(last_message).__name__}) - cannot route after agent"
+        )
+
+    return last_message
+
+
+def route_agent(
+    state: AgentGraphState,
+) -> list[str] | Literal[AgentGraphNode.AGENT, AgentGraphNode.TERMINATE]:
+    """Route after agent: handles all routing logic including control flow detection.
+
+    Routing logic:
+    1. If multiple tool calls exist, filter out control flow tools (EndExecution, RaiseError)
+    2. If control flow tool(s) remain, route to TERMINATE
+    3. If regular tool calls remain, route to specific tool nodes (return list of tool names)
+    4. If no tool calls, handle successive completions
+
+    Returns:
+        - list[str]: Tool node names for parallel execution
+        - AgentGraphNode.AGENT: For successive completions
+        - AgentGraphNode.TERMINATE: For control flow termination
+
+    Raises:
+        AgentNodeRoutingException: When encountering unexpected state (empty messages, non-AIMessage, or excessive completions)
+    """
+    messages = state.messages
+    last_message = __validate_last_message_is_AI(messages)
+
+    tool_calls = list(last_message.tool_calls) if last_message.tool_calls else []
+    tool_calls = __filter_control_flow_tool_calls(tool_calls)
+
+    if tool_calls and __has_control_flow_tool(tool_calls):
+        return AgentGraphNode.TERMINATE
+
+    if tool_calls:
+        return [tc["name"] for tc in tool_calls]
+
+    successive_completions = count_successive_completions(messages)
+
+    if successive_completions > MAX_SUCCESSIVE_COMPLETIONS:
+        raise AgentNodeRoutingException(
+            f"Agent exceeded successive completions limit without producing tool calls "
+            f"(completions: {successive_completions}, max: {MAX_SUCCESSIVE_COMPLETIONS}). "
+            f"This should not happen as tool_choice='required' is enforced at the limit."
+        )
+
+    if last_message.content:
+        return AgentGraphNode.AGENT
+
+    raise AgentNodeRoutingException(
+        f"Agent produced empty response without tool calls "
+        f"(completions: {successive_completions}, has_content: False)"
+    )

uipath_langchain/agent/react/terminate_node.py

@@ -0,0 +1,82 @@
+"""Termination node for the Agent graph."""
+
+from __future__ import annotations
+
+from typing import Any, NoReturn
+
+from langchain_core.messages import AIMessage
+from pydantic import BaseModel
+from uipath.agent.react import END_EXECUTION_TOOL, RAISE_ERROR_TOOL
+from uipath.runtime.errors import UiPathErrorCode
+
+from ..exceptions import (
+    AgentNodeRoutingException,
+    AgentTerminationException,
+)
+from .types import AgentGraphState, AgentTermination
+
+
+def _handle_end_execution(
+    args: dict[str, Any], response_schema: type[BaseModel] | None
+) -> dict[str, Any]:
+    """Handle LLM-initiated termination via END_EXECUTION_TOOL."""
+    output_schema = response_schema or END_EXECUTION_TOOL.args_schema
+    validated = output_schema.model_validate(args)
+    return validated.model_dump()
+
+
+def _handle_raise_error(args: dict[str, Any]) -> NoReturn:
+    """Handle LLM-initiated error via RAISE_ERROR_TOOL."""
+    error_message = args.get("message", "The LLM did not set the error message")
+    detail = args.get("details", "")
+    raise AgentTerminationException(
+        code=UiPathErrorCode.EXECUTION_ERROR,
+        title=error_message,
+        detail=detail,
+    )
+
+
+def _handle_agent_termination(termination: AgentTermination) -> NoReturn:
+    """Handle Command-based termination."""
+    raise AgentTerminationException(
+        code=UiPathErrorCode.EXECUTION_ERROR,
+        title=termination.title,
+        detail=termination.detail,
+    )
+
+
+def create_terminate_node(
+    response_schema: type[BaseModel] | None = None,
+):
+    """Handles Agent Graph termination for multiple sources and output or error propagation to Orchestrator.
+
+    Termination scenarios:
+    1. Command based termination with information in state (e.g: escalation)
+    2. LLM-initiated termination (END_EXECUTION_TOOL)
+    3. LLM-initiated error (RAISE_ERROR_TOOL)
+    """
+
+    def terminate_node(state: AgentGraphState):
+        if state.termination:
+            _handle_agent_termination(state.termination)
+
+        last_message = state.messages[-1]
+        if not isinstance(last_message, AIMessage):
+            raise AgentNodeRoutingException(
+                f"Expected last message to be AIMessage, got {type(last_message).__name__}"
+            )
+
+        for tool_call in last_message.tool_calls:
+            tool_name = tool_call["name"]
+
+            if tool_name == END_EXECUTION_TOOL.name:
+                return _handle_end_execution(tool_call["args"], response_schema)
+
+            if tool_name == RAISE_ERROR_TOOL.name:
+                _handle_raise_error(tool_call["args"])
+
+        raise AgentNodeRoutingException(
+            "No control flow tool call found in terminate node. Unexpected state."
+        )
+
+    return terminate_node

uipath_langchain/agent/react/tools/__init__.py

@@ -0,0 +1,7 @@
+from .tools import (
+    create_flow_control_tools,
+)
+
+__all__ = [
+    "create_flow_control_tools",
+]