The start of the holiday season in November did not stop us from deploying our largest slate of features and capabilities yet.

Codeium Live

You might remember our blog posts on context awareness earlier, and we realized that by building this, we had built infrastructure that could help anyone interact with the most up-to-date version of a repository, even outside of an IDE. By adding our indexing and retrieval logic on any codebase, we essentially solve the cutoff-date problem that ChatGPT and any LLM-only tool has, where any changes after the training time are not reflected in the response. To test the value of a more powerful version of ChatGPT for code-related questions, we built Codeium Live, a way to interact with the most up-to-date version of popular libraries and open source repositories, even those like LangChain that didn’t really make it any of the GPT model training runs. All in the browser. For free.

ChatGPT when asked about LangChain

Codeium Live when asked about LangChain

Expanded Full Repository Context Awareness Availability

On the topic of understanding the most up-to-date version of repositories, this parsing+embedding+indexing+retrieval logic became mature enough on Visual Studio Code for our SaaS users (both individual and Teams) that we have started to port it to our other environments. Namely, this month, we have added this logic to our SaaS JetBrains plugin.

Codeium Command

We continued to develop the in-IDE experience as well, launching Codeium Command, our prototype for allowing developers to provide instructive commands directly in-editor, rather than having to rely on a separate Chat panel. Ever found yourself writing a comment to instruct Autocomplete to do something, only to delete that comment afterwards since you never intended that to be checked into the codebase? Command is the solution to that. Currently available on Visual Studio Code.

Attribution

From the beginning, we have removed any non-permissive or copyright code from the training data of our proprietary LLMs. However, we knew due to the probabilistic nature of LLMs and the fact that we would still like to know if generated code was similar to permissively licensed code, we built and launched a post-generation attribution filter and audit log. Our post-generation attribution is much more robust than similar functionality from other code assistants, and our unique combination of proactive and reactive legal compliance will now give our enterprise customers the ultimate feeling of safety.

GPT-4 Experimentation

We’ve been experimenting with integrating GPT-4 into our system. We know our proprietary code-biased chat models are significantly faster than GPT-4 and believe that they are as good if not better than GPT-4 on the vast majority of code-related questions. But “vast majority” does not equate to “always,” and in the interest of always driving the theoretical best value to our users, we are experimenting with how to enable using GPT-4 in cases where it would be the best option. Thanks to all of you who have provided feedback so far.

The Mentions capability from the Chat panel (where you can use the @ symbol to reference pieces of the codebase to pull in as context) has now been expanded to @repo, where you can tag popular open-source libraries to chat with their code and documentation. These repo contexts are refreshed daily so you can always chat with the most recent version.

For example, you can tag @repo:material-ui to ask Codeium how to use a Material UI Button component:

Content

If you missed any of the content from or about Codeium in the last month, check them out!

With new surfaces, modalities, and features (not to mention an expansion of past capabilities), November unlocked a new speed for Codeium. We are excited to ride this momentum into 2024 - make sure you tell your friends and colleagues to join the party!

Changelist: November 2023