Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Inline completion strips out new line when it should be kept #1001

Open
krassowski opened this issue Sep 17, 2024 · 1 comment · May be fixed by #1002
Open

Inline completion strips out new line when it should be kept #1001

krassowski opened this issue Sep 17, 2024 · 1 comment · May be fixed by #1002
Labels
bug Something isn't working good first issue Good for newcomers

Comments

@krassowski
Copy link
Member

Description

New line is stripped out if model returns

```

prefix.

Reproduce

  1. Write comment like # load json file
  2. See suggestion gets applied like: # load json file import

Expected behavior

# load json file
import

Context

This is an issue in

jupyter-ai/packages/jupyter-ai-magics/jupyter_ai_magics/completion_utils.py

Lines 25 to 47 in fcb2d71

def post_process_suggestion(suggestion: str, request: InlineCompletionRequest) -> str:
"""Remove spurious fragments from the suggestion.
While most models (especially instruct and infill models do not require
any pre-processing, some models such as gpt-4 which only have chat APIs
may require removing spurious fragments. This function uses heuristics
and request data to remove such fragments.
"""
# gpt-4 tends to add "```python" or similar
language = request.language or "python"
markdown_identifiers = {"ipython": ["ipython", "python", "py"]}
bad_openings = [
f"```{identifier}"
for identifier in markdown_identifiers.get(language, [language])
] + ["```"]
for opening in bad_openings:
# ollama models tend to add spurious whitespace
if suggestion.lstrip().startswith(opening):
suggestion = suggestion.lstrip()[len(opening) :].lstrip()
# check for the prefix inclusion (only if there was a bad opening)
if suggestion.startswith(request.prefix):
suggestion = suggestion[len(request.prefix) :]
break

@krassowski krassowski added the bug Something isn't working label Sep 17, 2024
@krassowski
Copy link
Member Author

To fix it we need to:

- suggestion = suggestion.lstrip()[len(opening) :].lstrip() 
+ suggestion = suggestion.lstrip()[len(opening) :]
  # check for the prefix inclusion (only if there was a bad opening) 
- if suggestion.startswith(request.prefix): 
-      suggestion = suggestion[len(request.prefix) :]
+ if suggestion.lstrip().startswith(request.prefix): 
+      suggestion = suggestion.lstrip()[len(request.prefix) :]

Plus tests cases need to be added in

jupyter-ai/packages/jupyter-ai/jupyter_ai/tests/completions/test_handlers.py

Lines 101 to 133 in fcb2d71

expected_suggestions_cases = [
("```python\nTest python code\n```", "Test python code"),
("```\ntest\n```\n \n", "test"),
("```hello```world```", "hello```world"),
(" ```\nprint(test)\n```", "print(test)"),
("``` \nprint(test)\n```", "print(test)"),
]
@pytest.mark.parametrize(
"response,expected_suggestion",
expected_suggestions_cases,
)
async def test_handle_request_with_spurious_fragments(response, expected_suggestion):
inline_handler = MockCompletionHandler(
lm_provider=MockProvider,
lm_provider_params={
"model_id": "model",
"responses": [response],
},
)
dummy_request = InlineCompletionRequest(
number=1, prefix="", suffix="", mime="", stream=False
)
await inline_handler.handle_request(dummy_request)
# should write a single reply
assert len(inline_handler.messages) == 1
# reply should contain a single suggestion
suggestions = inline_handler.messages[0].list.items
assert len(suggestions) == 1
# the suggestion should include insert text from LLM without spurious fragments
assert suggestions[0].insertText == expected_suggestion

@krassowski krassowski added the good first issue Good for newcomers label Sep 17, 2024
@hiCXK hiCXK linked a pull request Sep 17, 2024 that will close this issue
Open
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
bug Something isn't working good first issue Good for newcomers
Projects
None yet
Milestone
No milestone
Development

Successfully merging a pull request may close this issue.

Fix inline completion bug, add test cases hiCXK/jupyter-ai
1 participant
@krassowski