DSLituiev
commented
1 year ago I seem getting a similar bug that the command is not being passed to the tool. Or sometimes not generated if "memory_prompts" kwarg is not provided
Setup
```python
from langchain.agents import AgentType
from langchain.chat_models import ChatOpenAI
from langchain.agents import initialize_agent
from langchain.prompts import MessagesPlaceholder
from langchain.memory import ConversationBufferMemory
from langchain.tools import BaseTool
from langchain.schema import BaseRetriever, Document
from langchain.tools import WikipediaQueryRun
from langchain.utilities import WikipediaAPIWrapper
from langchain.retrievers import WikipediaRetriever
llm_params = {
"engine": "gpt-35-turbo-16k",
"temperature": 0.5, # To avoid pure copy-pasting from docs lookup
"max_tokens": 6000
}
llm = ChatOpenAI(**llm_params)
wikipedia = WikipediaQueryRun(api_wrapper=WikipediaAPIWrapper())
class WikipediaSearchTool(BaseTool):
name = "wikipedia_search"
description = "Use this last. Useful for when you are not able to answer questions using the paper." + \
"Provide search terms consisting of a noun or noun phrase, like 'Canada' for canadian population"
def _run(self, query:str):
return Document(page_content=wikipedia.run(query))
async def _arun(self, query:str):
pass
tools = [WikipediaSearchTool(),]
chat_history = MessagesPlaceholder(variable_name="chat_history")
memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)
```
Agent not generating commands without memory_prompts argument
agent_chain = initialize_agent(
tools,
llm,
agent=AgentType.STRUCTURED_CHAT_ZERO_SHOT_REACT_DESCRIPTION,
verbose=True,
memory=memory,
agent_kwargs = {
"input_variables": ["input", "agent_scratchpad", ]
}
)
agent_chain.run(input="How many people live in Canada?")Output:
Failed to load default session, using empty session: HTTPConnectionPool(host='localhost', port=8000): Max retries exceeded with url: /sessions?name=default (Caused by NewConnectionError('<urllib3.connection.HTTPConnection object at ...>: Failed to establish a new connection: [Errno 61] Connection refused'))
> Entering new AgentExecutor chain...
Failed to persist run: HTTPConnectionPool(host='localhost', port=8000): Max retries exceeded with url: /chain-runs (Caused by NewConnectionError('<urllib3.connection.HTTPConnection object at ...>: Failed to establish a new connection: [Errno 61] Connection refused'))
I can find the answer to that using the Wikipedia search tool. Let me do that for you.
> Finished chain.
Out[125]: 'I can find the answer to that using the Wikipedia search tool. Let me do that for you.'Agent generating commands but not launching tools
agent_chain = initialize_agent(
tools,
llm,
agent=AgentType.STRUCTURED_CHAT_ZERO_SHOT_REACT_DESCRIPTION,
verbose=True,
memory=memory,
agent_kwargs = {
"memory_prompts": [chat_history],
"input_variables": ["input", "agent_scratchpad", "chat_history"]
}
)
agent_chain.run(input="How many people live in Canada?")Output:
Failed to load default session, using empty session: HTTPConnectionPool(host='localhost', port=8000): Max retries exceeded with url: /sessions?name=default (Caused by NewConnectionError('<urllib3.connection.HTTPConnection object at 0x179ce8b80>: Failed to establish a new connection: [Errno 61] Connection refused'))
> Entering new AgentExecutor chain...
Failed to persist run: HTTPConnectionPool(host='localhost', port=8000): Max retries exceeded with url: /chain-runs (Caused by NewConnectionError('<urllib3.connection.HTTPConnection object at 0x179ce9db0>: Failed to establish a new connection: [Errno 61] Connection refused'))
{
"action": "wikipedia_search",
"action_input": {
"query": {
"title": "Demographics of Canada"
}
}
}
> Finished chain.
Out[123]: '{\n "action": "wikipedia_search",\n "action_input": {\n "query": {\n "title": "Demographics of Canada"\n }\n }\n}'
clemlesne
dosubot[bot]
hdnh2006
System Info
I have been using STRUCTURED_CHAT_ZERO_SHOT_REACT_DESCRIPTION with a custom tool that takes 2 inputs, however have not been able to get the agent to produce the outputs we had before. Specifically, we're getting the intermediate, custom function input as output.
Ie instead of getting a value associated with a query "What is the
purchase totalfor customer 5432 on 07-03-2023?" We are now gettingThis did not occur before this weekend. Here's more code snippet in the below example
Who can help?
@hwcha
Information
Related Components
Reproduction
I also tried simplifying the agent call, etc to no avail:
conversational_agent = initialize_agent( agent=AgentType.STRUCTURED_CHAT_ZERO_SHOT_REACT_DESCRIPTION, tools=[database_tool], llm=llm, verbose=True, early_stopping_method='generate', memory=memory, SystemAgentPromptTemplate=prompt_template+"\n The only tool available is the database tool.", return_intermediate_steps=True, return_source_documents=True, handle_parsing_errors='Check your output and make sure there is an equal number of "{" and "}"' )
response = conversational_agent("What is the
purchase totalfor customer 5432 on 07-03-2023?") print(response['output'])