Introducing vector-triage
I just open-sourced vector-triage over the weekend:
It runs as a GitHub Action and helps detect duplicates and similar items for Issues and PRs.
The idea started from a Twitter discussion when Pieter (@steipete) asked about this kind of feature in another OSS project. That project needed extra third-party dependencies for vector DB and embeddings. I wanted a simpler path for GitHub-native repos, so I built this around GitHub Models, which are free for minimal usage.
What vector-triage is for
vector-triage is a GitHub Action that:
- embeds issue/PR content with GitHub Models
- searches a local SQLite index (vector + FTS)
- posts one managed triage comment when related items are found
It works for both Issues and Pull Requests.
Design goals
I kept the project focused on a few principles:
- zero-config setup
- deterministic behavior
- low comment noise
- safe defaults for OSS repos
Quick start
Setup is a single workflow file:
name: Triage
on:
issues:
types: [opened, edited]
pull_request_target:
types: [opened, synchronize]
permissions:
contents: write
issues: write
pull-requests: write
models: read
jobs:
triage:
runs-on: ubuntu-latest
concurrency:
group: triage-index
cancel-in-progress: false
steps:
- uses: rizwankce/[email protected]That is the full setup.
How it behaves
- First run: it creates the index and does not post noise comments.
- Later runs: if matches exist, it posts or updates one managed triage comment.
- If no matches are found, it stays quiet.
- Recoverable failures are logged as warnings and exit non-fatally.
Built with Codex
I built most of this in a weekend with the help of Codex.
The biggest win was speed of iteration: shaping the retrieval behavior, comment format, and safety model quickly, then hardening with tests and E2E workflow checks.
Links
- Repository: https://github.com/rizwankce/vector-triage
- Action usage:
rizwankce/[email protected]