OpenReview: code review with vectors

OpenReview is one of a growing set of tools trying to turn code review from a generic chatbot pass into a repository-aware workflow that runs inside pull requests. (github.com) The project is open source, self-hosted, and currently marked beta in its public repository, which had about 1,200 stars and 88 forks on GitHub when it was crawled in April 2026. (github.com) In plain terms, code review software checks proposed changes before they are merged, and newer systems use large language models to read diffs, inspect files, and draft comments. OpenReview says developers can trigger a review by mentioning `@openreview` in a pull request comment after connecting a GitHub App. (github.com) The software runs in an isolated Vercel Sandbox, clones the pull request branch, installs dependencies, and can run linters, formatters, and tests before posting inline suggestions. Its public workflow diagram also says it can commit and push fixes for formatting, lint errors, and some simple bugs. (github.com) That design targets a common complaint about automated review bots: they often comment on the changed lines without enough context from the rest of the codebase. Research on automated review has increasingly focused on context-aware defect localization rather than snippet-only generation or comment style. (openreview.net) The basic idea behind vector search is to turn code or documents into number lists that preserve meaning, then retrieve nearby matches instead of exact keyword hits. Microsoft’s documentation for vector search describes embeddings as vectors stored in an index so systems can return semantically similar results at query time. (github.com) OpenReview’s public repository does not spell out a standalone vector database in the README, but it does emphasize “full repo access,” custom review skills, and agent-driven exploration of files and command output during review runs. That means the project’s repository awareness, at least in the current public version, appears to come from live codebase access and workflow orchestration rather than a separately documented retrieval layer. (github.com) The project also sits inside a broader push by developer-tool companies to move artificial intelligence systems from chat windows into software pipelines. Vercel says OpenReview was built internally to test its own stack, including Vercel Workflow, Vercel Sandbox, the AI Software Development Kit, and Next.js route handlers. (github.com; github.com) For engineering teams, the practical question is less whether a bot can write comments than whether it can find the right files, run the right checks, and leave comments specific enough to act on. OpenReview’s beta label and self-hosted setup suggest Vercel is still treating that as an infrastructure problem as much as a language-model problem. (github.com)