Merge pull request #304 from langchain-ai/nc/15apr/tool-calls

Upgrade to new tool calling api

backend/app/agent.py

@@ -8,8 +8,7 @@
 )
 from langgraph.checkpoint import CheckpointAt
 
-from app.agent_types.google_agent import get_google_agent_executor
-from app.agent_types.openai_agent import get_openai_agent_executor
+from app.agent_types.tools_agent import get_tools_agent_executor
 from app.agent_types.xml_agent import get_xml_agent_executor
 from app.chatbot import get_chatbot_executor
 from app.checkpoint import PostgresCheckpoint
@@ -80,22 +79,22 @@ def get_agent_executor(
 ):
     if agent == AgentType.GPT_35_TURBO:
         llm = get_openai_llm()
-        return get_openai_agent_executor(
+        return get_tools_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(
+        return get_tools_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(
+        return get_tools_agent_executor(
             tools, llm, system_message, interrupt_before_action, CHECKPOINTER
         )
     elif agent == AgentType.CLAUDE2:
         llm = get_anthropic_llm()
-        return get_xml_agent_executor(
+        return get_tools_agent_executor(
             tools, llm, system_message, interrupt_before_action, CHECKPOINTER
         )
     elif agent == AgentType.BEDROCK_CLAUDE2:
@@ -105,12 +104,12 @@ def get_agent_executor(
         )
     elif agent == AgentType.GEMINI:
         llm = get_google_llm()
-        return get_google_agent_executor(
+        return get_tools_agent_executor(
             tools, llm, system_message, interrupt_before_action, CHECKPOINTER
         )
     elif agent == AgentType.OLLAMA:
         llm = get_ollama_llm()
-        return get_openai_agent_executor(
+        return get_tools_agent_executor(
             tools, llm, system_message, interrupt_before_action, CHECKPOINTER
         )
 

backend/app/agent_types/openai_agent.py → backend/app/agent_types/tools_agent.py

@@ -1,9 +1,14 @@
-import json
+from typing import cast
 
 from langchain.tools import BaseTool
-from langchain.tools.render import format_tool_to_openai_tool
 from langchain_core.language_models.base import LanguageModelLike
-from langchain_core.messages import SystemMessage, ToolMessage
+from langchain_core.messages import (
+    AIMessage,
+    FunctionMessage,
+    HumanMessage,
+    SystemMessage,
+    ToolMessage,
+)
 from langgraph.checkpoint import BaseCheckpointSaver
 from langgraph.graph import END
 from langgraph.graph.message import MessageGraph
@@ -12,7 +17,7 @@
 from app.message_types import LiberalToolMessage
 
 
-def get_openai_agent_executor(
+def get_tools_agent_executor(
     tools: list[BaseTool],
     llm: LanguageModelLike,
     system_message: str,
@@ -27,13 +32,16 @@ async def _get_messages(messages):
                 _dict["content"] = str(_dict["content"])
                 m_c = ToolMessage(**_dict)
                 msgs.append(m_c)
+            elif isinstance(m, FunctionMessage):
+                # anthropic doesn't like function messages
+                msgs.append(HumanMessage(content=str(m.content)))
             else:
                 msgs.append(m)
 
         return [SystemMessage(content=system_message)] + msgs
 
     if tools:
-        llm_with_tools = llm.bind(tools=[format_tool_to_openai_tool(t) for t in tools])
+        llm_with_tools = llm.bind_tools(tools)
     else:
         llm_with_tools = llm
     agent = _get_messages | llm_with_tools
@@ -43,7 +51,7 @@ async def _get_messages(messages):
     def should_continue(messages):
         last_message = messages[-1]
         # If there is no function call, then we finish
-        if "tool_calls" not in last_message.additional_kwargs:
+        if not last_message.tool_calls:
             return "end"
         # Otherwise if there is, we continue
         else:
@@ -54,16 +62,13 @@ async def call_tool(messages):
         actions: list[ToolInvocation] = []
         # Based on the continue condition
         # we know the last message involves a function call
-        last_message = messages[-1]
-        for tool_call in last_message.additional_kwargs["tool_calls"]:
-            function = tool_call["function"]
-            function_name = function["name"]
-            _tool_input = json.loads(function["arguments"] or "{}")
-            # We construct an ToolInvocation from the function_call
+        last_message = cast(AIMessage, messages[-1])
+        for tool_call in last_message.tool_calls:
+            # We construct a ToolInvocation from the function_call
             actions.append(
                 ToolInvocation(
-                    tool=function_name,
-                    tool_input=_tool_input,
+                    tool=tool_call["name"],
+                    tool_input=tool_call["args"],
                 )
             )
         # We call the tool_executor and get back a response
@@ -72,12 +77,10 @@ async def call_tool(messages):
         tool_messages = [
             LiberalToolMessage(
                 tool_call_id=tool_call["id"],
+                name=tool_call["name"],
                 content=response,
-                additional_kwargs={"name": tool_call["function"]["name"]},
-            )
-            for tool_call, response in zip(
-                last_message.additional_kwargs["tool_calls"], responses
             )
+            for tool_call, response in zip(last_message.tool_calls, responses)
         ]
         return tool_messages
 

backend/app/checkpoint.py

@@ -2,13 +2,22 @@
 from datetime import datetime
 from typing import AsyncIterator, Optional
 
+from langchain_core.messages import BaseMessage
 from langchain_core.runnables import ConfigurableFieldSpec, RunnableConfig
 from langgraph.checkpoint import BaseCheckpointSaver
 from langgraph.checkpoint.base import Checkpoint, CheckpointThreadTs, CheckpointTuple
 
 from app.lifespan import get_pg_pool
 
 
+def loads(value: bytes) -> Checkpoint:
+    loaded: Checkpoint = pickle.loads(value)
+    for key, value in loaded["channel_values"].items():
+        if isinstance(value, list) and all(isinstance(v, BaseMessage) for v in value):
+            loaded["channel_values"][key] = [v.__class__(**v.__dict__) for v in value]
+    return loaded
+
+
 class PostgresCheckpoint(BaseCheckpointSaver):
     class Config:
         arbitrary_types_allowed = True
@@ -47,7 +56,7 @@ async def alist(self, config: RunnableConfig) -> AsyncIterator[CheckpointTuple]:
                             "thread_ts": value[1],
                         }
                     },
-                    pickle.loads(value[0]),
+                    loads(value[0]),
                     {
                         "configurable": {
                             "thread_id": thread_id,
@@ -70,7 +79,7 @@ async def aget_tuple(self, config: RunnableConfig) -> Optional[CheckpointTuple]:
                 ):
                     return CheckpointTuple(
                         config,
-                        pickle.loads(value[0]),
+                        loads(value[0]),
                         {
                             "configurable": {
                                 "thread_id": thread_id,
@@ -92,7 +101,7 @@ async def aget_tuple(self, config: RunnableConfig) -> Optional[CheckpointTuple]:
                                 "thread_ts": value[1],
                             }
                         },
-                        pickle.loads(value[0]),
+                        loads(value[0]),
                         {
                             "configurable": {
                                 "thread_id": thread_id,

backend/app/llms.py

@@ -5,10 +5,11 @@
 
 import boto3
 import httpx
-from langchain_community.chat_models import BedrockChat, ChatAnthropic, ChatFireworks
+from langchain_community.chat_models import BedrockChat, ChatFireworks
 from langchain_community.chat_models.ollama import ChatOllama
 from langchain_google_vertexai import ChatVertexAI
 from langchain_openai import AzureChatOpenAI, ChatOpenAI
+from langchain_anthropic import ChatAnthropic
 
 logger = logging.getLogger(__name__)
 
@@ -67,7 +68,11 @@ def get_anthropic_llm(bedrock: bool = False):
         )
         model = BedrockChat(model_id="anthropic.claude-v2", client=client)
     else:
-        model = ChatAnthropic(temperature=0, max_tokens_to_sample=2000)
+        model = ChatAnthropic(
+            model_name="claude-3-haiku-20240307",
+            max_tokens_to_sample=2000,
+            temperature=0,
+        )
     return model