initialize_agent with zero_shot_react_description, talks to itself, produces conversational buffer memory issues

[mroytman83]

I'm running the executor off a flask backend, and when the answer from the llm to my number is POSTED, strangely, the agent begins to talk to itself, with seemingly no human message. I store all the messages within the executor's memory, and this is the error that it throws:

2023-08-17 14:05:05.122663
+122*******(my number)

Entering new AgentExecutor chain...
The user is initiating a conversation. No specific question or request has been made.
Action: None needed at this point.
Final Answer: Hello! How can I assist you with your fitness and wellness goals today?

Finished chain.
6.702334880828857
127.******* - - [17/Aug/2023 09:05:11] "POST /sms HTTP/1.1" 200 -
2023-08-17 14:05:12.594520
+187*******(twilio number)
2023-08-17 14:05:12.608324
+187*******(twilio number)

Entering new AgentExecutor chain...

Entering new AgentExecutor chain...
2023-08-17 14:05:14.522510
+187*******(twilio number)

Entering new AgentExecutor chain...
There is no question provided for me to answer. <- no message is passed
Action: None
Final Answer: Could you please provide a question?

Finished chain.
[2023-08-17 09:05:16,692] ERROR in app: Exception on /sms [POST]
Traceback (most recent call last):
File "/opt/anaconda3/lib/python3.9/site-packages/flask/app.py", line 2190, in wsgi_app
response = self.full_dispatch_request()
File "/opt/anaconda3/lib/python3.9/site-packages/flask/app.py", line 1486, in full_dispatch_request
rv = self.handle_user_exception(e)
File "/opt/anaconda3/lib/python3.9/site-packages/flask/app.py", line 1484, in full_dispatch_request
rv = self.dispatch_request()
File "/opt/anaconda3/lib/python3.9/site-packages/flask/app.py", line 1469, in dispatch_request
return self.ensure_sync(self.view_functions[rule.endpoint])(**view_args)
File "/Users/michaelroytman/Desktop/bolic-ai/app.py", line 53, in inbound_sms
msg_cache=ai.calc(client_number, body, user_dict)
File "/Users/michaelroytman/Desktop/bolic-ai/service/bolicai.py", line 264, in calc
bolicai_response = user_dict[client_number[1:]].respond_to(body, client_number)
File "/Users/michaelroytman/Desktop/bolic-ai/service/bolicai.py", line 168, in respond_to
response=agent_chain.run(input=text)
File "/opt/anaconda3/lib/python3.9/site-packages/langchain/chains/base.py", line 441, in run
return self(kwargs, callbacks=callbacks, tags=tags, metadata=metadata)[
File "/opt/anaconda3/lib/python3.9/site-packages/langchain/chains/base.py", line 245, in call
final_outputs: Dict[str, Any] = self.prep_outputs(
File "/opt/anaconda3/lib/python3.9/site-packages/langchain/chains/base.py", line 339, in prep_outputs
self.memory.save_context(inputs, outputs)
File "/opt/anaconda3/lib/python3.9/site-packages/langchain/memory/chat_memory.py", line 37, in save_context
self.chat_memory.add_user_message(input_str)
File "/opt/anaconda3/lib/python3.9/site-packages/langchain/schema/memory.py", line 100, in add_user_message
self.add_message(HumanMessage(content=message))
File "/opt/anaconda3/lib/python3.9/site-packages/langchain/load/serializable.py", line 74, in init
super().init(**kwargs)
File "pydantic/main.py", line 341, in pydantic.main.BaseModel.init
pydantic.error_wrappers.ValidationError: 1 validation error for HumanMessage
content

[dosubot[bot]]

🤖

Hello,

From the error logs you've provided, it seems like the issue is related to the LangChain's memory management when dealing with user messages. Specifically, the error is thrown when trying to validate the HumanMessage content.

The error message pydantic.error_wrappers.ValidationError: 1 validation error for HumanMessage content suggests that the content of the HumanMessage is not passing the validation checks set by Pydantic, a data validation library used by LangChain.

This could be due to several reasons:

1. The content of the HumanMessage is None or empty. This could happen if there's an issue with how the user messages are being fetched or parsed in your application. Please ensure that the user messages are being correctly fetched and passed to the LangChain's AgentExecutor.

2. The content of the HumanMessage is not of the expected type. The HumanMessage content should be a string. If it's of a different type, the validation would fail. Please ensure that the user messages are of the correct type before they are passed to the AgentExecutor.

3. There's an issue with the LangChain's memory management. If the AgentExecutor is not correctly managing the memory, it could lead to issues when trying to save the context of the conversation. Please ensure that the AgentExecutor is correctly initialized and that it's correctly managing the memory.

To debug this issue, you could add logging statements in your application to log the content of the user messages before they are passed to the AgentExecutor. This would help you identify if the issue is with the user messages or with the AgentExecutor.

If the issue persists, please provide more information about how you're initializing the AgentExecutor and how you're fetching and passing the user messages to it. This would help us better understand the issue and provide a more accurate solution.

Best,
Dosu

Sources

• libs/langchain/langchain/agents/agent.py
• libs/langchain/langchain/tools/base.py
• libs/langchain/langchain/chains/base.py

---

This response is meant to be useful and save you time. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

If you want to continue the conversation, start your reply with @dosu-bot.