Gradient AI: tinyurl.com/gradient-ai
Get the code: github.com/gkamradt/RAGWithGr...
Get updates from me: mail.gregkamradt.com/
Greg’s Info:
- Twitter: / gregkamradt
- Newsletter: mail.gregkamradt.com/
- Website: gregkamradt.com/
- LinkedIn: / gregkamradt
- Work with me: tiny.one/TEi2HhN
- Contact Me: Twitter DM, LinkedIn Message, or contact@dataindependent.com
🔥 as usual Greg! Thanks!
Love it thanks
Any guidance for RAG with complex documents like Engineering Documents that have a lot of figures (think of an IKEA manual with instructions) and then have instructions further along that refer to those figures. For example: if the compressor is being extra noisy check the air filter by taking part A in figure 3 and unscreweing the 4 bolts (C 8’ figure), etc
ya, I would check out level 3 of kzhead.info/sun/a7ODc5Zpi2SJf2w/bejne.html which talks about different representations of raw data. You'll need to do some serious post processing on your chunks to make sure the data can be referenced correctly
I think something like llamaparse can deal with many figures and formulae
Simple and useful
Super simple
Love this! Easy RAG! The best! So much easier.
As prompt content sizes continue to increase, will we get to a point where RAG won’t be needed? Eventually will we be to fit everything in the prompt?
eh, you won't be able to fit all of wikipedia into a prompt, so we'll need to select somehow. My take is that data selection (retrieval) will be a thing for a while
I just saw a study that showed Claude and chatGPT are very inaccurate when it comes to using long content added in the context window
@@DataIndependent thank you for your feedback
@@HashimWarren interesting, would be a good experiment to compare both approaches
@@HashimWarren yeah Greg himself did a "needle in the haystack" thing. LLMs do not have the capability yet to reliably retrieve facts and reason over them in a large context window. Also you pay by token so retrieval is very much relevant still and will probably continue to do so.
nice
Bye!