LlamaIndex Sessions: Practical Tips and Tricks for Productionizing RAG (feat. Sisil @ Jasper)
ฝัง
- เผยแพร่เมื่อ 17 ก.พ. 2024
- In this video, Sisil Mehta (ML eng @, Jasper) walks through practical tips and tricks that his team implemented for productionizing a RAG system at Jasper.ai, backed by LlamaIndex abstractions.
These tricks include the following:
1. Picking a proper PDF parser that can maintain semantic structure, parse text from tables/images, and be represented as XML or Markdown
2. Adding the right "layers" of metadata; besides global document context, also inject summary context from "sub-documents" to more precisely localize context.
3. Hybrid fusion between different retrieval methods
4. LLM-powered reranking. Reduce token usage by reranking summaries that reference underlying chunks.
5. Use XML and emotion prompting to get well-structured outputs free of hallucinations
Super comprehensive. Thanks for this.
This was an excellent talk. Thanks so much for sharing your experience and this RAG framework. If there could be a follow up sometime with a sample notebook that uses these techniques, and a code walkthrough video , I’m sure many people would greatly benefit from it.
Sisil's presentation was exemplary, addressing key pain points with innovation.
Congratulations on your work!. We as a company also trying to solve all pain points you have mentioned in the pdf docs area.
Thanks to Jerry for spotlighting talent like Sisil. Excited for more content!
a showcase sample notebook would be deeply appreciated!
Would have loved if you shared the slides in the description. :)
amazing!!! Sisil really explained the difference between benchmarks and real world openended questions!
btw, could you include the datasets name Sisil talked about regarding "retrieval" and "reranking" on evaluation?
Is there a sample code you can post a link to, specifically for indexing the subdocs, then the chunks and retrieving them?
this was fantastic, can you provide a tutorial notebook?
how does lexical indexing work for subdocs
I wonder what is the PDF parse Sisil 's team is using?
Adobe Extract PDF API or something. I haven't tested it but it works pretty well from what I've heard.
There is an opensource version as well (Works in Linux Distro) called unstructured. You just need to do pip install unstructured[all].
I didn't catch what was the pdf parser used? Can you name it please?
I think it was adobe api