langchain-ai · nfcampos · Feb 5, 2024 · Feb 4, 2024
diff --git a/backend/app/agent.py b/backend/app/agent.py
@@ -49,25 +49,38 @@ def get_agent_executor(
     tools: list,
     agent: AgentType,
     system_message: str,
+    interrupt_before_action: bool,
 ):
     if agent == AgentType.GPT_35_TURBO:
         llm = get_openai_llm()
-        return get_openai_agent_executor(tools, llm, system_message, CHECKPOINTER)
+        return get_openai_agent_executor(
+            tools, llm, system_message, interrupt_before_action, CHECKPOINTER
+        )
     elif agent == AgentType.GPT_4:
         llm = get_openai_llm(gpt_4=True)
-        return get_openai_agent_executor(tools, llm, system_message, CHECKPOINTER)
+        return get_openai_agent_executor(
+            tools, llm, system_message, interrupt_before_action, CHECKPOINTER
+        )
     elif agent == AgentType.AZURE_OPENAI:
         llm = get_openai_llm(azure=True)
-        return get_openai_agent_executor(tools, llm, system_message, CHECKPOINTER)
+        return get_openai_agent_executor(
+            tools, llm, system_message, interrupt_before_action, CHECKPOINTER
+        )
     elif agent == AgentType.CLAUDE2:
         llm = get_anthropic_llm()
-        return get_xml_agent_executor(tools, llm, system_message, CHECKPOINTER)
+        return get_xml_agent_executor(
+            tools, llm, system_message, interrupt_before_action, CHECKPOINTER
+        )
     elif agent == AgentType.BEDROCK_CLAUDE2:
         llm = get_anthropic_llm(bedrock=True)
-        return get_xml_agent_executor(tools, llm, system_message, CHECKPOINTER)
+        return get_xml_agent_executor(
+            tools, llm, system_message, interrupt_before_action, CHECKPOINTER
+        )
     elif agent == AgentType.GEMINI:
         llm = get_google_llm()
-        return get_google_agent_executor(tools, llm, system_message, CHECKPOINTER)
+        return get_google_agent_executor(
+            tools, llm, system_message, interrupt_before_action, CHECKPOINTER
+        )
     else:
         raise ValueError("Unexpected agent type")
 
@@ -77,6 +90,7 @@ class ConfigurableAgent(RunnableBinding):
     agent: AgentType
     system_message: str = DEFAULT_SYSTEM_MESSAGE
     retrieval_description: str = RETRIEVAL_DESCRIPTION
+    interrupt_before_action: bool = False
     assistant_id: Optional[str] = None
     user_id: Optional[str] = None
 
@@ -88,6 +102,7 @@ def __init__(
         system_message: str = DEFAULT_SYSTEM_MESSAGE,
         assistant_id: Optional[str] = None,
         retrieval_description: str = RETRIEVAL_DESCRIPTION,
+        interrupt_before_action: bool = False,
         kwargs: Optional[Mapping[str, Any]] = None,
         config: Optional[Mapping[str, Any]] = None,
         **others: Any,
@@ -107,7 +122,9 @@ def __init__(
                     _tools.extend(_returned_tools)
                 else:
                     _tools.append(_returned_tools)
-        _agent = get_agent_executor(_tools, agent, system_message)
+        _agent = get_agent_executor(
+            _tools, agent, system_message, interrupt_before_action
+        )
         agent_executor = _agent.with_config({"recursion_limit": 50})
         super().__init__(
             tools=tools,
@@ -257,6 +274,11 @@ def __init__(
     .configurable_fields(
         agent=ConfigurableField(id="agent_type", name="Agent Type"),
         system_message=ConfigurableField(id="system_message", name="Instructions"),
+        interrupt_before_action=ConfigurableField(
+            id="interrupt_before_action",
+            name="Tool Confirmation",
+            description="If Yes, you'll be prompted to continue before each tool is executed.\nIf No, tools will be executed automatically by the agent.",
+        ),
         assistant_id=ConfigurableField(
             id="assistant_id", name="Assistant ID", is_shared=True
         ),

diff --git a/backend/app/agent_types/google_agent.py b/backend/app/agent_types/google_agent.py
@@ -16,6 +16,7 @@ def get_google_agent_executor(
     tools: list[BaseTool],
     llm: LanguageModelLike,
     system_message: str,
+    interrupt_before_action: bool,
     checkpoint: BaseCheckpointSaver,
 ):
     def _get_messages(messages):
@@ -105,4 +106,6 @@ async def call_tool(messages):
     # This compiles it into a LangChain Runnable,
     # meaning you can use it as you would any other runnable
     app = workflow.compile(checkpointer=checkpoint)
+    if interrupt_before_action:
+        app.interrupt = ["action:inbox"]
     return app
diff --git a/backend/app/agent_types/openai_agent.py b/backend/app/agent_types/openai_agent.py
@@ -16,6 +16,7 @@ def get_openai_agent_executor(
     tools: list[BaseTool],
     llm: LanguageModelLike,
     system_message: str,
+    interrupt_before_action: bool,
     checkpoint: BaseCheckpointSaver,
 ):
     async def _get_messages(messages):
@@ -119,4 +120,6 @@ async def call_tool(messages):
     # This compiles it into a LangChain Runnable,
     # meaning you can use it as you would any other runnable
     app = workflow.compile(checkpointer=checkpoint)
+    if interrupt_before_action:
+        app.interrupt = ["action:inbox"]
     return app
diff --git a/backend/app/agent_types/xml_agent.py b/backend/app/agent_types/xml_agent.py
@@ -63,6 +63,7 @@ def get_xml_agent_executor(
     tools: list[BaseTool],
     llm: LanguageModelLike,
     system_message: str,
+    interrupt_before_action: bool,
     checkpoint: BaseCheckpointSaver,
 ):
     formatted_system_message = xml_template.format(
@@ -153,4 +154,6 @@ async def call_tool(messages):
     # This compiles it into a LangChain Runnable,
     # meaning you can use it as you would any other runnable
     app = workflow.compile(checkpointer=checkpoint)
+    if interrupt_before_action:
+        app.interrupt = ["action:inbox"]
     return app
diff --git a/backend/app/storage.py b/backend/app/storage.py
@@ -144,7 +144,7 @@ def get_thread(user_id: str, thread_id: str) -> Thread | None:
 def get_thread_messages(user_id: str, thread_id: str):
     """Get all messages for a thread."""
     config = {"configurable": {"user_id": user_id, "thread_id": thread_id}}
-    app = get_agent_executor([], AgentType.GPT_35_TURBO, "")
+    app = get_agent_executor([], AgentType.GPT_35_TURBO, "", False)
     checkpoint = app.checkpointer.get(config) or empty_checkpoint()
     with ChannelsManager(app.channels, checkpoint) as channels:
         return {
@@ -158,7 +158,7 @@ def get_thread_messages(user_id: str, thread_id: str):
 def post_thread_messages(user_id: str, thread_id: str, messages: Sequence[AnyMessage]):
     """Add messages to a thread."""
     config = {"configurable": {"user_id": user_id, "thread_id": thread_id}}
-    app = get_agent_executor([], AgentType.GPT_35_TURBO, "")
+    app = get_agent_executor([], AgentType.GPT_35_TURBO, "", False)
     checkpoint = app.checkpointer.get(config) or empty_checkpoint()
     with ChannelsManager(app.channels, checkpoint) as channels:
         channel = channels[MESSAGES_CHANNEL_NAME]

diff --git a/frontend/src/components/Config.tsx b/frontend/src/components/Config.tsx
@@ -63,13 +63,18 @@ function Types(props: {
   );
 }
 
-function Label(props: { id?: string; title: string }) {
+function Label(props: { id?: string; title: string; description?: string }) {
   return (
     <label
       htmlFor={props.id}
-      className="block font-medium leading-6 text-gray-400 mb-2"
+      className="flex flex-col font-medium leading-6 text-gray-400 mb-2"
     >
-      {props.title}
+      <div>{props.title}</div>
+      {props.description && (
+        <div className="font-normal text-sm text-gray-600 whitespace-pre-line">
+          {props.description}
+        </div>
+      )}
     </label>
   );
 }
@@ -84,7 +89,11 @@ function StringField(props: {
 }) {
   return (
     <div>
-      <Label id={props.id} title={props.title} />
+      <Label
+        id={props.id}
+        title={props.title}
+        description={props.field.description}
+      />
       <textarea
         rows={4}
         name={props.id}
@@ -109,7 +118,11 @@ export default function SingleOptionField(props: {
 }) {
   return (
     <div>
-      <Label id={props.id} title={props.field.title} />
+      <Label
+        id={props.id}
+        title={props.field.title}
+        description={props.field.description}
+      />
       <fieldset>
         <legend className="sr-only">{props.field.title}</legend>
         <div className="space-y-2">
@@ -170,7 +183,11 @@ function MultiOptionField(props: {
 }) {
   return (
     <fieldset>
-      <Label id={props.id} title={props.title ?? props.field.items?.title} />
+      <Label
+        id={props.id}
+        title={props.title ?? props.field.items?.title}
+        description={props.field.description}
+      />
       <div className="space-y-2">
         {orderBy(props.field.items?.enum)?.map((option) => (
           <div className="relative flex items-start" key={option}>
@@ -285,6 +302,7 @@ function fileId(file: File) {
 const ORDER = [
   "system_message",
   "retrieval_description",
+  "interrupt_before_action",
   "tools",
   "llm_type",
   "agent_type",
@@ -483,6 +501,30 @@ export function Config(props: {
                 readonly={readonly}
               />
             );
+          } else if (value.type === "boolean") {
+            return (
+              <SingleOptionField
+                key={key}
+                id={key}
+                field={{
+                  ...value,
+                  type: "string",
+                  enum: ["Yes", "No"],
+                }}
+                title={title}
+                value={values?.configurable?.[key] ? "Yes" : "No"}
+                setValue={(value: string) =>
+                  setValues({
+                    ...values,
+                    configurable: {
+                      ...values!.configurable,
+                      [key]: value === "Yes",
+                    },
+                  })
+                }
+                readonly={readonly}
+              />
+            );
           } else if (
             value.type === "array" &&
             value.items?.type === "string" &&