Data collection and vectorization have always been major pain points in RAG application development. Commercial solutions work smoothly but come with a high price tag, while building your own stack from scratch is incredibly time-consuming. The open-source project Supavec aims to solve this problem. Positioned as an open-source alternative to Carbon.ai, it claims you can get a production-ready vector search and chat API deployed in under 5 minutes.


It currently holds 630 upvotes on Product Hunt and 620 stars on GitHub, so it's been gaining solid traction already. The project is built with Next.js, Supabase, Tailwind, Bun and Upstash, uses a multi-tenant architecture, and implements data isolation via Row Level Security (RLS).

## Performance Metrics

The development team has published several key performance indicators:

- Batch embedding processing cuts OpenAI costs by 65%

- Configurable chunk size and overlap parameters improve retrieval recall by 12 percentage points

- Hybrid filtering (file_id + cosine similarity) delivers a P95 latency of 210ms

- Supports both streaming and standard responses, selectable per request

These numbers aren't groundbreaking, but they're solid and practical for an open-source project.

## Pricing Model

It offers a free tier with 100 requests per month, a Basic plan with 750 requests, and an Enterprise plan with 5,000 requests. All plans are pay-as-you-go, so the barrier to entry is very low.

## The Companion Scraper: DeepScrape

Many people point out that RAG applications need more than just vectorization and retrieval—they require a reliable crawling system. That's where DeepScrape comes in; the name is straightforward—it's built for deep web scraping.

This project is more interesting than it sounds: it's not just a basic web scraper, it also comes with LLM-powered content extraction. It uses Playwright for browser automation and pairs it with GPT-4o for accurate content extraction. Its key features include:

- Fit-Markdown extraction: Filters out noise like navigation bars and footers via link density scoring, generating clean, LLM-ready markdown

- Deterministic CSS extraction: No LLM required, pure CSS selectors, zero hallucinations, zero extra cost

- LLM-powered structured extraction: Converts raw web content into validated JSON, with schema validation and token-aware chunking

- Self-healing extraction: LLMs infer and cache CSS selectors, and automatically re-infer them after website updates

- Change tracking: Git-style diffing to show exactly what content changed on a page

It also has a particularly useful feature: it can turn any website into an MCP endpoint that AI agents can call directly.

## Combining the Two Projects

Used together, these two projects cover the entire workflow for a complete RAG application: DeepScrape converts web content into clean markdown or structured data, while Supavec handles vectorization, storage, and retrieval.

The core idea is clear: open-source, self-hosted, full control over your data. If you're tired of the limitations of commercial APIs or want to cut costs, this combo is definitely worth trying.

## Parting Thoughts

The growing number of projects like this shows that RAG infrastructure is rapidly becoming commoditized. What used to require a dedicated engineering team to build can now be assembled from a handful of open-source projects. That's great news for users, but it also means competition in this space will only get fiercer.

Supavec has gained good early traction, but it hasn't seen large-scale production validation yet. DeepScrape is full-featured, but its documentation is quite extensive. Whether you should adopt this stack depends entirely on your specific use case.

Either way, the tools are out there—whether you use them is up to you.