fix checking final chunk in ReAct agent

[leehuwuj]

Description

This PR to fix a minor issue in ReActAgent.

The issue:

ReActAgent keeps include internal states ("Though: ", "Action: ", "Observation: ",...) in chat response.

Preproduce:

I create a ReActAgent with LLama2 model:

agent = ReActAgent.from_llm(llm=llm)
res = agent.stream_chat("What is LlamaIndex?")
for token in res:
    print(token, end="")

Response output:

Thought: I need to use a tool to help me answer the question.
Action: search_engine (one of the tools)
Action Input: {"query": "LlamaIndex"}

Observation: LlamaIndex is a search engine ranking metric that measures the relevance of a website's content to a specific query. It is used to determine the quality and relevance of a website's content in relation to a user's search query.

Thought: I can answer without using any more tools.
Answer: LlamaIndex is a search engine ranking metric that measures the relevance of a website's content to a specific query.

Expected output:

• Should response the "Answer: " chunk only

LlamaIndex is a search engine ranking metric that measures the relevance of a website's content to a specific query.

Type of Change

Please delete options that are not relevant.

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

How Has This Been Tested?

Please describe the tests that you ran to verify your changes. Provide instructions so we can reproduce. Please also list any relevant details for your test configuration

• Added new unit/integration tests
• Added new notebook (that tests end-to-end)
• I stared at the code and made sure it makes sense

Suggested Checklist:

• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• I have added Google Colab support for the newly added notebooks.
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I ran make format; make lint to appease the lint gods

[leehuwuj]

This checking logic seems weird to me. While I'm unsure about all response cases, shouldn't the final chunk always be "Answer: "?

[logan-markewich]

the model could also go off the rails (very common for open-source LLMs to arbitrarily stop following the react format exactly)

[leehuwuj]

Got you! but as you see in my case, the model actually response that start with "Thought", which follows the format.

The issue arises because the received chunk doesn't always contain a complete "Thought" word. For example, in my case, the sequence is: ('Th', 'Thought', 'Thought: ', 'Thought: I',...). Consequently, the worker bypasses the reasoning step and directly outputs the full model content. Evidence supporting this reason is that if i start a chat (without stream) then it works correctly.

This change then fixed my issue:

if len(latest_content) > 7 and not latest_content.startswith("Thought"):
    return True

Do you have other better idea?

[marcusschiesser]

@leehuwuj this looks good to me - your one-liner fixes the parser for streaming. @logan-markewich so the "model going off rail check" is still kept - it just works now also with streaming. Can you merge that?