LlamaIndex Webinar: Improving RAG with Advanced Parsing + Metadata Extraction

In this video we cohost a workshop with the cofounders of Deasie (Reece, Leonard, Mikko) on improving RAG with advanced parsing and metadata.
​The data processing layer is one of the most important pieces to get right for RAG. This means that AI engineers need to make careful decisions in terms of parsing and transformations - including metadata extraction and chunking - in order to make sure that their e2e QA system is surfacing relevant results.
​This is a nice two-part workshop that demonstrates the following:
- ​the value of good parsing itself over complex documents, with LlamaParse
- ​the value of additional value of adding in metadata through Deasie's powerful automated labeling platform
​We show overall experimental results over research papers validating the combination of parsing + metadata for good performance.

[@SamiSabirIdrissi]

Overall i think this is super dope! I can’t wait to try this. The Increase routing accuracy capability is wild. Pulling relevant data with high accuracy is extremely important! 💪⚡️

[@SamiSabirIdrissi]

Very very interesting, i feel like this is similar to Azure’s document intelligence feature?

[@awakenwithoutcoffee]

yup, similar to unstructured / Llamaparse + LlamaExtract (new) . There is also new OCR models like ColPali!

Same for the metadata tags generation: another open AI GPT wrapper doing generation or mapping depending if the tags are suggested or not. As shown, the best result is obtained with the custom metadata. It means that humans are still in need to do the most difficult and time-consuming task, i.e. defining the custom tasks...😢

[@awakenwithoutcoffee]

nah I can't see this not being automated in the foreseeable future. There are already OCR models with long context memory that are able to create metadata tags. Give it a few months and this metadata "problem" will be solved.

[@pin65371]

It seems to me like this would get much more effective with a graph system? When Jerry was asking about how the data would be retrieved it seemed like graph would work well with this. When you ask a question the LLM first would retrieve relevant parent metadata. From there it can branch out from there. The advantage with that would be that connections that maybe arent so obvious with vector would be very obvious with graph. Also with graph at least you have visibility to be able to manually go in and see what is going on. I liked that last question as well. It seems like maybe it wasnt something they thought about but they might look at how to implement something like that. Tokens are getting so cheap now that it would make sense. Especially if lets say you are using the openai 4o-mini model its 30 cents for a million tokens output. Just getting it to output some extra metadata would essentially be free and would only make the whole system more efficient in the long run.

[@awakenwithoutcoffee]

I agree but it still too expensive and difficult to fully automate correctly. Let's keep in touch trough to the comments as us engineers are looking to for production ready techniques. My take is that graphRAG is not ready yet but it might be early next year (for enterprise).

The example with PyPDF is not correct as nobody is using PyPDF texts extracted per page. Instead, there is post-processing on the raw text. All these startups founders think that we are dummies and propose in their "products" the recipes that we are all using for months or years without pretending to build a company on the top of them. Same for metadata.... Almost nothing new here. 😢😮

[@awakenwithoutcoffee]

you bring up an important point: The part about cross-page context confused me since Jerry basically didn't know why this was happening. Have you found additional information or techniques yourself ? I'm looking for production ready techniques for meta-data extraction. One alternative new approach is ColPali.

[@MatijaGrcic]

This was great, thanks for sharing.

[@isle1009]

8:16 Does Deasie support languages other than English well, especially Korean?

[@Deasie]

Yes, we do support other languages, including Korean!