nkaretnikov

Are you asking about the query/tokenization or the overall architecture?

I assume it's the former. It's explained in the post after the diagram, see "The diagram illustrates a series of steps [...]".

We get a user query. Based on that query, we pull relevant parts from the doc. Then, we submit both the query and the doc parts to an LLM. This limits the amount of data we need to send and allows the user to know which parts of the doc are relevant.

My colleagues at Quansight released Ragna. It's pretty cool!

In the blogpost, there's a demo: you upload a pdf, ask ML model a question, it answers and shows which pages it looked at in the document. If you don't trust the output, you can double-check the doc since you have the page numbers.

Ragna currently ships with pre-built extensions for OpenAI, MosaicML, Anthropic, and local LLMs as well as the Chroma and LanceDB vector databases. And it's open source, too!

I've written an article about PyTorch! It's meant for non-PyTorch devs and focuses on software architecture. Enjoy!