Relace (YC W23) Launches Fast and Reliable Models for Code Generation and Retrieval

Hello HN community! We’re Preston and Eitan, cofounders of Relace (https://relace.ai), and we're excited to introduce our platform. Our mission is to make building code generation agents both easier and more affordable. The Challenges of Building Reliable Code Agents Creating effective code generation agents involves more than just developing simple prototypes. Two main issues arise once you move beyond initial stages: how do you reliably apply code changes, and how do you effectively manage codebase context? Solving These Problems Economically and Efficiently Relace is designed to address these challenges at a fraction of the cost and latency compared to existing solutions. Our first model, released in February, is the Fast Apply model. This model merges code snippets into files at a speed of 4300 tokens per second, making it more reliable in terms of merge errors than models like Sonnet, Qwen, and Llama. Each file typically processes in about 900 milliseconds, providing a seamless user experience and saving approximately 40% on Claude 4 output tokens. Enhancing Code Retrieval Our second model focuses on retrieval. By accurately identifying and retrieving only the files relevant to a user's request, it saves significantly on state-of-the-art input token costs and minimizes unnecessary file viewing by code agents. Our reranker can scan a million-line codebase in just 1 to 2 seconds, and our embedding model outperforms any other for retrieval, as tested on TypeScript/React repositories. Foundational Issues in Coding Agents While there are various approaches to building coding agents, the ability to edit code reliably and retrieve the most pertinent parts of a codebase will continue to be fundamental. We are thrilled to contribute to making these capabilities accessible to millions of developers who might otherwise find the costs prohibitive. Real-World Usage and Integration These models are already being used in production, processing millions of requests each week. If you’ve interacted with platforms like Lovable, Create.xyz, Magic Patterns, Codebuff, or Tempo Labs, you have likely benefited from Relace’s technology. Try Relace Today We’ve made free access available for prototyping on our website, and the usage limits should suffice for personal coding and small projects. Should you encounter any issues or find the limits insufficient, please let us know. Additionally, Relace seamlessly integrates with open-source IDEs like Continue.dev. We encourage you to try out Relace at https://app.relace.ai. Our documentation is comprehensive and can be found at https://docs.relace.ai. We value your feedback and are eager to hear how you find our tools. Thank you for your support, and happy coding!