noahshinn / reflexion-draft Goto Github PK

thanks a lot for your great work. i am interested in this test very much.
i have issues that need your help:

i have tried this test but failed again and again.
there are so many environment issues, such as : python version, textworld version, alfworld version, etc...

is it possible that you provide a step by step guide that let me make it work.

several issues that i met:

1. infos = textworld.EnvInfos(won=True, admissible_commands=True, expert_type=expert_type, expert_plan=expert_plan, extras=["gamefile"])
   returns error: expert type is not there. it seems that the alfworld and text world version are not compatible. then i removed these two parameters to walk around...

my version is as below, not sure is it is correct:
textworld 1.5.3
alfworld 0.2.2

2. cann't find "domain: '$ALFWORLD_DATA/logic/alfred.pddl'"
   on my pc, the file is at: "/Users/xyz/Library/Python/3.7/lib/python/site-packages/alfworld/data/alfred.pddl"
   i hardcoded the file to walk around the issue. but it seems that the installation of alfred is not correct on my side...

3. cannot find : challenge_few_shot_examples.txt
   i replaced it with : reflexion_few_shot_examples.txt

4. other issues that i do not mentioned by details here...

You said

To replicate results from the ReAct study (Yao et al., 2023), we prompted the agent with six few-shot examples of successful trajectories.

Does this mean you are doing prompt engineering, using example shown in 4.3 Reflexion Enables More Intuitive Search Queries?

Put these prompt before user input, and just get result?

But I think this is not the case, because AFAIK, to get this kind of result, you have to use prompt like https://github.com/hwchase17/langchain/blob/b7ebb8fe3009dd791b562968524718e20bfb4df8/langchain/agents/conversational/prompt.py#L14-L27 it usually need to include you Must use this format