RAG-in-a-box with tenant isolation and usage meters included

What it does

Supavec is an open-source RAG-as-a-Service platform pitched as an alternative to Carbon.ai. It exposes REST endpoints for document ingestion and chat-style retrieval, backed by Supabase for storage and vector search. The stack includes Next.js, Tailwind, Bun, and Upstash Redis for rate limiting.

The interesting bit

Most open-source RAG demos stop at a Jupyter notebook; this one ships row-level security for team isolation, usage-based billing tiers, and batch embedding pipelines that claim a 65 % cut in OpenAI costs. It also offers streaming or standard API responses and a visual embedding debugger in the chat UI—nice touches for a project that is essentially plumbing.

Key highlights

• Multi-tenant by default: Supabase RLS keeps teams’ data separated.
• Cost controls: batch embeddings and configurable chunk/overlap settings that reportedly lift recall by 12 points.
• Performance targets: hybrid filtering by file_id plus cosine similarity hits P95 in 210 ms.
• Observability built-in: PostHog analytics, request-level tracing IDs, and async usage logging.
• Developer experience: sliding-window Redis rate limiting, progressive onboarding, and a live embedding preview.

Caveats

• The README still advertises “620⭐ on GitHub” and Product Hunt stats that appear stale next to the repo’s current star count.
• The “any data source” promise is vague; the docs shown center on file-based ingestion via a REST API.
• API documentation currently links to a personal fork (taishikato/supavec), suggesting the repository structure may still be shifting.

Verdict

Good for teams that want a full-stack RAG service with billing and multi-tenancy wired up. Less ideal if you need a lightweight library or support for data sources beyond the file-ingestion API the README actually describes.