Anthropic published this in September last year, but I only came across this while following this blog post. I’m still working through the blog post, but I can’t help but to feel increasingly bullish about the Cloudflare Developer Platform.

Relatedly, a friend also shared this blog post by Shortwave (which is even more dated), where they get into the technical details of how they implement their AI assistant. These RAG pipelines can really get arbitrarily complex.

It’s hard to make sense of this phenomenon in the RAG space. On the one hand, there are managed services that abstract away the complexities of ingesting sources and retrievals from these sources:

• Graphlit: a friend uses it and raves about how it has existing integration with many sources. However, he says that the costs are very high. I can’t help but it abstracts away a little too much, but I can see how this could be a quick way to spin up an MVP.
• AutoRAG by Cloudflare: I came across a YouTube video that showed how the RAG system was set up in <5 minutes. It certainly seems to offer more control than Graphlit and I have faith that Cloudflare pricing would be reasonable.

On the other hand, there are companies that are literally just doing document parsing, e.g. Reducto.

• They raised a Series A led by Benchmark, who are certainly no fools. (Not to mention, there are many open-source libraries that promise to do the same thing.)
• A friend who has tried it (at some point) said it was no better than Sonnet 3.5.

The age of AI is coming, but I would much rather be building end-user products and services, rather the tooling around them. I look forward to getting my hands dirty and trying these out, but, from a “making sense of the industry” perspective, I can’t help feeling I’ll at best be a blind man groping one part of the elephant.