Datapizza AI Framework

Hey PH 👋

After a year shipping RAG systems and AI agents for real customers, we kept hitting the same wall: heavy abstractions made debugging painful, slowed onboarding, and hid the parts we needed to control. We didn’t want another “magic” layer—we wanted a thin, modular framework that stays close to provider SDKs and lets engineers intervene at any depth.

So we focussed on:
- Low abstraction, full control: no black boxes; every step is observable (tracing, structured logs, clear error handling).
- Composable by design: swap or customize any module—chunking, embeddings, indexing, retrieval, re-ranking, caching, observability, evaluation—without fighting the framework.
- Fast for engineers: if you know OpenAI/Anthropic SDKs, you’re productive in minutes. Granular deps keep deployments lean.
- No lock-in: open-source, provider-agnostic, fits your stack.

Before launch, we ran Datapizza AI in 10+ enterprise projects: faster builds, simpler debugging, surgical customization where it mattered.

That it’s built for the messy reality of production, not just demos. And that we’re releasing it openly—so teams everywhere can build reliable GenAI without wrestling over-complex abstractions.

Quick-start, full RAG examples, and multi-agent templates are in the repo. Kick the tires, break things, tell us what to improve.

Love the thin, modular approach. Tracing, structured logs, clear error handling, and swap-anything modules are exactly what production teams need. Open source, provider agnostic, close to SDKs, and already battle tested. Excited to try it on our next RAG. Congrats!

Let's goooo Datapizza 🚀🚀

Hi Product Hunt,
I am kinda excited to finally launch our framework.

When I joined Datapizza, we found most existing implementation too black boxes, we wanted to build something from scratch and our customization, like, our R&D experimentation were just sitting in scripts with no clear path to production.

Thus we created something fully modular, but only with the abstraction that were truly needed to keep it as lightweight as we could.

Like, bridging the gap between research and production-ready code was brutal, but as for now, I think it was all worth.

Even if now I'm focused more on R&D, it was a whole ride to see a year of work finally open sourced.

Hope this may also help the community as it helped us internally.

Feels great to finally share what we’ve been working on.

We started building Datapizza AI because the existing tools we used never felt quite right: too many layers, too little control, and no real way to reuse what our R&D team kept discovering (customization).

We wanted a framework that stays close to the SDKs, fully modular and transparent/observable, where every component — from RAG to agents and evaluation — can evolve an be highly customized in real world use cases.

Turning months of research into something usable and production-ready was one the hardest parts, but also the most rewarding.

After a year of iteration and real-world testing, it’s finally open source and I still can't believe it.

Can’t wait to see what the community builds with it. 🚀

The Datapizza framework is an incredible piece of work. We're using it within our AI outreach and experimentation group, also to help less nerdy people understand what an industrialized approach to GenAI means, with surprising results.

It feels amazing to finally open source something we’ve been building and refining for over a year.

What’s even better is seeing the community already extending it, adapting it to their needs, and using it in their companies to build GenAI solutions that are truly production-ready.

Our goal has always been simple: make GenAI create real impact for people and organizations. Seeing what's happening out there makes me feel we’re on the right track 🔥