Fix: improve file filtering, fix yaml parsing, add openrouter support.#71
Fix: improve file filtering, fix yaml parsing, add openrouter support.#71redreamality wants to merge 0 commit intoThe-Pocket:mainfrom
Conversation
utils/crawl_local_files.py
Outdated
| relpath = os.path.relpath(filepath, directory) | ||
| else: | ||
| relpath = filepath | ||
| rel_root = os.path.relpath(root, directory) if use_relative_paths else root |
There was a problem hiding this comment.
This should have already been addressed by another PR?
#67
Could you check if that PR made such optimization?
utils/fix_yaml.py
Outdated
| @@ -0,0 +1,20 @@ | |||
| import re | |||
|
|
|||
| def add_indentation(text): | |||
There was a problem hiding this comment.
I would prefer to not fix indentation as a post processing.
If an LLM failed with outputting yaml, it could be caused by many reasons. I'm bit worried that fixing problematic one will still get a result that's off in contnet. I would just let LLM retry.
There was a problem hiding this comment.
I agree that fixing indent for yaml is a temporal solution. In fact, this need occurs from use case for large repos. A new issue to address this might be more suitable.
FYI.
I tested it on some inside projects, on deepseek, gemini, qwen, and claude.
when total tokens below 100k, things usually work well.
when ~700k tokens, for gemini-2.5, sometimes it successes after retry; for gemini-2.0, it nearly fails all the time (at same point with indent problem).
nodes.py
Outdated
|
|
||
| {language_instruction}Analyze the codebase context. | ||
| Identify the top 5-10 core most important abstractions to help those new to the codebase. | ||
| Identify the top 5-20 core most important abstractions to help those new to the codebase. |
There was a problem hiding this comment.
Do you find 20 is a good change? I tune this before and find abstractions above 10 become less interesting and a bit low level to read. Maybe make this a tunable parameters?
Love it! Could you check if it has been implemented by previous PR?
I'd rather just let LLM retry. My experience is that, when LLM messes up the indentation, it usually also messes up the content a bit.
Could you make it tunable?
Love them! |
|
now addressed in #74 |
2 participants
crawl_local_files.pywith folder-level exclusion, partially solve Can it support large repo? #23fix_yaml.pyutility for YAML indentation fixesnodes.pyto support up to 20 core abstractions Can it support large repo? #23