Skip to main content

​
Introduction

Relace Repos is a fully managed version control system specifically designed for AI applications, offering:
  • Lightweight read/write operations enabling easy integration with your backend.
  • Agentic search that maintains low latency on large codebases.
  • Multi-tenancy that scales to millions of repositories across thousands of users.
  • Permissive rate limits that allow interactions across many repos simultaneously.
  • Full git compatibility, allowing easy interactions from the git CLI.

​
Source Management for AI

Building a scalable AI code editing application usually requires:
  • Durable storage of source code
  • Versioning, so that unsuccessful or undesired changes can be easily reverted
  • Low-latency reads/writes on the working version of the code
  • Support for automated interactions with thousands of isolated repos
  • Integration with GitHub, to allow human developers to easily collaborate on the same code base
Relace Repos offers a centralized platform that satisfies all of these requirements, seamlessly integrated with our state-of-the-art models and agents.

​
Existing Alternatives

Industry standard version control services like GitHub are designed primarily for human developers, which leads to some pain points for AI applications.

​
Service limits

Humans have low frequency manual interactions through websites and the git CLI. This results in relatively low limits that are insufficient for large scale automated AI applications:
  • A single account/organization may not exceed 100,000 repos
  • REST API requests may not exceed 5,000 per hour (15,000 for a GitHub app owned by an enterprise organization)
Most text-to-app systems treat each user application as a single repo, which leads to the repo limit being reached very quickly. Systems with many concurrent users will also hit the rate limit relatively quickly. Assuming that a single AI edit requires at least 2 API calls (pulling and pushing), your capacity would be ~2 requests per second.

​
Integration complexity

Human developer workflows are less constrained and more complex than a typical AI application. This leads to a system that requires many distinct steps to make a simple change to the code base:
  1. Clone/checkout the repository locally
  2. Edit the files
  3. Stage and commit the changes
  4. Push to the remote
Since commits are made locally, you must setup a git library or the git CLI on your host. Given that most AI workflows are run in a serverless or sandboxed environment, this also means that full source retrieval must happen every time you spin up the environment. This contributes to high cold-start latency, and often necessitates some sort of file caching strategy. A relatively simple workflow where an agent reads a code base and edits some files can quickly become a highly complex infrastructure problem. Providing the right code context to your LLM is essential to produce the best results without sacrificing cost or latency. Every Relace Repo can be searched with our Fast Agentic Search agent: call the search endpoint with a query in natural language, and the agent explores the repo and reports back the relevant files with line ranges and an explanation.

​
Next Steps

For more details on how to set up Relace Repos as your agents source control system see our onboarding guide.