Reduce the number of LLM parsing error with conversational chats

[lesser-panda]

A lot of times, when the LLM responses a Final Answer with markdown in the answer text (e.g. asking the LLM to write some code), the parser raises an OutputParserException("Could not parse LLM output: ..."), when the answer from the LLM is totally acceptable.

This change should reduce the number of such occurances.

• Description: Reduce the number of LLM parsing error with conversational chats,
• Issue: the issue ValueError: Could not parse LLM output: #1358 ,
• Dependencies: None,
• Tag maintainer: @hinthornw @baskaryan,

For example, this response from gpt-turbo-3.5 for "How to convert a list of dict to a JSON in Python" seems totally acceptable as the final answer, but will raise the OutputParserException, even with handle_parsing_errors=True. This PR fixes it.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Ignored Deployment

Name	Status	Preview	Comments	Updated (UTC)
langchain	⬜️ Ignored (Inspect)	Visit Preview		Jul 31, 2023 1:05am

[lesser-panda]

I understand this is probably not the most elegant solution, so any suggestions are welcomed!

[R0315]

I'm currently facing this parsing issue problem and i tried appling this solution but i'm just getting the below output rather than the actual answer. I asked it to write a python function that adds two numbers together (which I was doing as a test to see if the bot worked and kept getting the parsing error).

{ "action": "Final Answer", "action_input": "Sure! Here's a Python function that adds two numbers together:\n\npython\ndef add_numbers(a, b):\n return a + b\n" }

[nfcampos]

Thanks, no longer relevant, I suggest using an agent with tool-calling