Open Source Ecosystems

NaLCoS (NAtural Language COmmit Search) is a command-line tool for searching commit messages in your repository in natural language.

The key features are:

Search commit messages in both local and remote GitHub repositories.
Search for commits in a specific branch.
Look back any number of commits in history.
Open the commit link directly in your browser if retrieved from GitHub.

Internally, NaLCoS uses Sentence Transformers with pre-trained weights from multi-qa-MiniLM-L6-cos-v1. I chose this particular model because it has a good Performance vs Speed tradeoff. Since this model was designed for semantic search and has been pre-trained on 215M (question, answer) pairs from diverse sources, it is a good choice for tasks such as finding similarity between two sentences.

NaLCoS encodes the query string and all the commits into their corresponding vector embeddings and computes the cosine similarity between the query and all the commits. This is then used to rank the commits.

Why did I build this?

Most of the times when I've used Machine Learning till now, has been in dedicated environments such as Google Colab or Kaggle. I had been learning Natural Language Processing for a while and wanted to use transformers to build something different that is not very resource (read GPU) intensive and can be used like an everyday tool.

Though many Transformer models are far from fitting this description, I found that distilled models are not as hungry as their older siblings are infamous for. Searching for Git commits using natural language was something on which I could not find any pre-existing tool and thus decided to give this a shot.

Though there are various improvements left, I'm happy with what this initially turned out to be. I'm eager to see what further enhancements can be made to this to make it more efficient and useful.

Requirements

NaLCoS uses the following packages:

Sentence Transformers for the Transformer model.
Git Python for local Git operations.
GitHub API for interacting with GitHub repos.
Rich for well-formatted CLI output.

Installation

Installing with `pip` (Recommended)

Install with pip or your favourite PyPi manager:

$ pip install nalcos

Run NaLCoS on a repository of your choice. For example:

$ nalcos "handle nan issues" "numpy/numpy" --github

To see all available options, run with the --help flag:

$ nalcos --help

Note: When you run the nalcos command for the first time, it will, download the model which would be cached and used the next time you run NaLCoS.

Installing bleeding edge from the GitHub repository

Clone the repository:

$ git clone https://github.com/thepushkarp/nalcos.git

This also downloads the model weights stored in the nalcos/models directory so you don't have to download them while running the model for the first time.

cd into the nalcos directory:

$ cd nalcos

Create a virtual environment (click here to read about activating virtualenv):

$ virtualenv venv

Activate virtualenv (for Linux and MacOS):

$ source ./venv/bin/activate

Activate virtualenv (for Windows):

$ cd venv/Scripts/
$ activate

Install the requirements and the module:

$ pip install -r requirements.txt
$ pip install -e .

Run NaLCoS on a repository of your choice. For example:

$ nalcos "handle nan issues" "numpy/numpy" --github

To see all available options, run with the --help flag:

$ nalcos --help

Usage

A detailed information about the usage of NaLCoS can be found below:

usage: nalcos [-h] [-g] [-n N_MATCHES] [-b BRANCH] [-l LOOK_PAST] [-s] [-v] [--version] query location

Search a commit in your git repository using natural language.

positional arguments:
  query                 The query to search for similar commit messages.
  location              The repository path to search in. If '-g' or '--github' flag is not passed, searches
                        locally in the path specified, else takes in a remote GitHub repository name in the
                        format '{owner}/{repo_name}'

optional arguments:
  -h, --help            show this help message and exit
  -g, --github          Search on GitHub instead of searching in a local repository. Due to API limits
                        currently this allows for around 15 lookups per hour from your IP.
  -n N_MATCHES, --n-matches N_MATCHES
                        The number of matching results to return. Default 10.
  -b BRANCH, --branch BRANCH
                        The branch to search in. If not specified, the current branch will be used by default.
  -l LOOK_PAST, --look-past LOOK_PAST
                        Look back this many commits. Default 100.
  -s, --show-score      Shows the Cosine similarity score between the query and the retrieved commit messages.
                        1 is the best match and -1 is the worst.
  -v, --verbose         Show the entire commit message and not just the commit title.
  --version             show program's version number and exit

Examples

Input:

$ nalcos "handle nan issues" "numpy/numpy" --github

Output:

Found 100 commits.

                                                  Commits related to "handle nan issues" in "numpy/numpy"

 No.  Commit ID  Commit Message                                                                             Commit Author        Commit Date          

  1.  b6d7c4680  BUG: Fixed an issue wherein certain `nan<x>` functions could fail for object arrays        Bas van Beek         2021-09-03T13:41:54Z 
  2.  e4f85b08c  Merge pull request #19863 from BvB93/nanquantile                                           Charles Harris       2021-09-13T23:21:51Z 
  3.  ecba7133f  MAINT: Let `_remove_nan_1d` attempt to identify nan-containing object arrays               Bas van Beek         2021-09-05T21:46:34Z 
  4.  95e5d5abb  BUG: Fixed an issue wherein `nanpercentile` and `nanquantile` would ignore the dtype for   Bas van Beek         2021-09-11T11:54:56Z 
                 all-nan arrays                                                                                                                       
  5.  b3a66e88b  Merge pull request #19821 from BvB93/nanfunctions                                          Charles Harris       2021-09-05T23:32:30Z 
  6.  dc7dafe70  Merge pull request #19869 from mhvk/median_scalar_nan                                      Charles Harris       2021-09-14T21:09:26Z 
  7.  9ef778330  TST: Add more tests for `nanmedian`, `nanquantile` and `nanpercentile`                     Bas van Beek         2021-09-03T15:01:57Z 
  8.  6ba48721e  BUG: ensure np.median does not drop subclass for NaN result.                               Marten van Kerkwijk  2021-09-13T19:50:54Z 
  9.  e62aa4968  Merge pull request #19854 from BvB93/nanfunctions                                          Charles Harris       2021-09-09T15:14:09Z 
 10.  268e8e885  TST: Make nanfunc test ignore overflow instead of xfailing test                            Sebastian Berg       2021-09-07T22:55:41Z

Future plans

Please visit the NaLCoS To Do Project Board to see current status and future plans.

Known issues

Not all retrieved results are always relevant. I could think of two primary reasons for this:

The data the model was pre-trained on is not representative of how people write commit messages. Since commit messages usually contain technical jargon, merge commit messages, abbreviations and other non-common terms, the model (which has a limited vocabulary) is not able to generalize well to this data.
Two commits may be related even when their commit messages may not be similar and similarly two commit messages maybe unrelated even when their commit messages are similar. We often need more metadata (such as lines changes, files changed) etc. to make the predictions more accurate.

Contributing

Any suggestions, improvements or bug reports are welcome.

If you want to discuss any aspect of the project, please use the Discussions Tab.
You can submit your idea by opening an issue or creating a Pull Request.
If you'd like to improve the code, make sure you stick to the existing code style and naming conventions.