Improve

Overview¶

The improve tool scans the PR code changes, and automatically generates suggestions for improving the PR code. The tool can be triggered automatically every time a new PR is opened, or it can be invoked manually by commenting on any PR:

/improve

Example usage¶

Manual triggering¶

Invoke the tool manually by commenting /improve on any PR. The code suggestions by default are presented as a single comment:

To edit configurations related to the improve tool, use the following template:

/improve --pr_code_suggestions.some_config1=... --pr_code_suggestions.some_config2=...

For example, you can choose to present the suggestions as commitable code comments, by running the following command:

/improve --pr_code_suggestions.commitable_code_suggestions=true

Note that a single comment has a significantly smaller PR footprint. We recommend this mode for most cases. Also note that collapsible are not supported in Bitbucket. Hence, the suggestions are presented there as code comments.

Automatic triggering¶

To run the improve automatically when a PR is opened, define in a configuration file:

[github_app]
pr_commands = [
    "/improve",
    ...
]

[pr_code_suggestions]
num_code_suggestions_per_chunk = ...
...

The pr_commands lists commands that will be executed automatically when a PR is opened.
The [pr_code_suggestions] section contains the configurations for the improve tool you want to edit (if any)

Extended mode¶

An extended mode, which does not involve PR Compression and provides more comprehensive suggestions, can be invoked by commenting on any PR by setting:

[pr_code_suggestions]
auto_extended_mode=true

(This mode is true by default).

Note that the extended mode divides the PR code changes into chunks, up to the token limits, where each chunk is handled separately (might use multiple calls to GPT-4 for large PRs). Hence, the total number of suggestions is proportional to the number of chunks, i.e., the size of the PR.

Configuration options¶

General options

num_code_suggestions	Number of code suggestions provided by the 'improve' tool. Default is 4 for CLI, 0 for auto tools.
extra_instructions	Optional extra instructions to the tool. For example: "focus on the changes in the file X. Ignore change in ...".
rank_suggestions	If set to true, the tool will rank the suggestions, based on importance. Default is false.
commitable_code_suggestions	If set to true, the tool will display the suggestions as commitable code comments. Default is false.
persistent_comment	If set to true, the improve comment will be persistent, meaning that every new improve request will edit the previous one. Default is false.
enable_help_text	If set to true, the tool will display a help text in the comment. Default is true.

params for 'extended' mode

auto_extended_mode	Enable extended mode automatically (no need for the --extended option). Default is true.
num_code_suggestions_per_chunk	Number of code suggestions provided by the 'improve' tool, per chunk. Default is 5.
rank_extended_suggestions	If set to true, the tool will rank the suggestions, based on importance. Default is true.
max_number_of_calls	Maximum number of chunks. Default is 5.
final_clip_factor	Factor to remove suggestions with low confidence. Default is 0.9.

Usage Tips¶

Extra instructions

Extra instructions are very important for the imrpove tool, since they enable you to guide the model to suggestions that are more relevant to the specific needs of the project.

Be specific, clear, and concise in the instructions. With extra instructions, you are the prompter. Specify relevant aspects that you want the model to focus on.

Examples for extra instructions:

[pr_code_suggestions] # /improve #
extra_instructions="""\
Emphasize the following aspects:
- Does the code logic cover relevant edge cases?
- Is the code logic clear and easy to understand?
- Is the code logic efficient?
...
"""

Use triple quotes to write multi-line instructions. Use bullet points to make the instructions more readable.

Review vs. Improve tools comparison

The review tool includes a section called 'Possible issues', that also provide feedback on the PR Code. In this section, the model is instructed to focus only on major bugs and issues.
The improve tool, on the other hand, has a broader mandate, and in addition to bugs and issues, it can also give suggestions for improving code quality and making the code more efficient, readable, and maintainable (see here).
Hence, if you are interested only in feedback about clear bugs, the review tool might suffice. If you want a more detailed feedback, including broader suggestions for improving the PR code, also enable the improve tool to run on each PR.

A note on code suggestions quality¶

While the current AI for code is getting better and better (GPT-4), it's not flawless. Not all the suggestions will be perfect, and a user should not accept all of them automatically. Critical reading and judgment are required.
While mistakes of the AI are rare but can happen, a real benefit from the suggestions of the improve (and review) tool is to catch, with high probability, mistakes or bugs done by the PR author, when they happen. So, it's a good practice to spend the needed ~30-60 seconds to review the suggestions, even if not all of them are always relevant.
The hierarchical structure of the suggestions is designed to help the user to quickly understand them, and to decide which ones are relevant and which are not:
- Only if the Category header is relevant, the user should move to the summarized suggestion description
- Only if the summarized suggestion description is relevant, the user should click on the collapsible, to read the full suggestion description with a code preview example.

In addition, we recommend to use the exra_instructions field to guide the model to suggestions that are more relevant to the specific needs of the project.
Consider also trying the Custom Suggestions Tool 💎, that will only propose suggestions that follow specific guidelines defined by user.