Codeium’s proprietary context engine builds a deep understanding of your codebase.

Historically, code-generation approaches focused on fine-tuning large language models (LLMs) on a codebase, which is difficult to scale to the needs of every individual user. A more recent and popular approach leverages retrieval-augmented generation (RAG), which focuses on techniques to construct highly relevant, context-rich prompts to elicit accurate answers from an LLM.

The Codeium team has taken an extremely thoughtful and optimized RAG approach to codebase context, and has seen great success in producing high-quality suggestions and few hallucinations. This applies across the board to Autocompete, Chat, and Command.

Codeium offers full fine-tuning for enterprises, and the best solution combines fine-tuning with RAG. Nonetheless, Codeium’s RAG approach is highly effective at codebase context personalization on its own.

​
Default Context

Out of the box, Codeium takes multiple relevant sources of context into consideration.

  • The current file and other open files in your IDE, which are often very relevant to the code you are currently writing.
  • The entire local codebase is then indexed (including files that are not open), and relevant code snippets are sourced by Codeium’s retrieval engine as you write code, ask questions, or invoke commands.
  • For Pro users, we offer expanded context lengths increased indexing limits, and higher limits on custom context and pinned context items.
  • For Teams and Enterprise users, Codeium can also index remote repositories. This is useful for companies whose development organization works across multiple repositories.

​
Context Pinning

Developers have the option to offer additional guidance by pinning custom context. You can find this option under the context tab of the chat panel.

You can choose to pin directories, files, repositories, or code context items (functions, classes, etc.) as persistent context. Models reference these items for every suggestion, across Autocomplete, Chat, and Command.

​
Best Practices

Context Pinning is great when your task in your current file depends on information from other files. Try to pin only what you need. Pinning too much may slow down or negatively impact model performance.

Here are some ideas for effective context pinning:

  • Module Definitions: pinning class/struct definition files that are inside your repo but in a module separate from your currently active file.
  • Internal Frameworks/Libraries: pinning directories with code examples for using frameworks/libraries.
  • Specific Tasks: pinning a file or folder defining a particular interface (e.g., .proto files, abstract class files, config templates).
  • Current Focus Area: pinning the “lowest common denominator” directory containing the majority of files needed for your current coding session.
  • Testing: pinning a particular file with the class you are writing unit tests for.

​
Chat-Specific Context Features

When conversing with Codeium Chat, you have various ways of leveraging codebase context, like @-mentions or custom guidelines. See the Chat page for more information.

​
Frequently Asked Questions (FAQs)

​
Does Codeium index my codebase?

Yes, Codeium does index your codebase. It also uses LLMs to perform retrieval-augmented generation (RAG) on your codebase using techniques we invented like M-Query.

Indexing performance and features vary based on your workflow and your Codeium plan. For more information, please visit our context awareness page.

OverviewCommon Issues