hey Lars! good question so for PDF-type documents, RAG is always going to be the way to go. there's all this context within the pdf itself - paragraph after paragraph - you can't put all that into a regular DB (i mean, you can, but you shouldn't lol - for cost and scale) so RAG helps with these types of documents because it gives you/llm a way to search within the PDF's so for example, say my documents are annual reports, and i'm looking for "overall revenue growth over past year for Ford". if you have a db schema, you could probably find the doc for Ford, then get the raw pdf, feed it into the llm, and ask it to answer. RAG essentially does that by giving you a retrieval/search method - so you query your knowledge base with keyword or your whole question, and you get back chunks from the document - which then is much easier to feed to the llm. my example was pretty basic, but this becomes really powerful when there's questions that have answers across multiple documents - it can give you chunks from multiple documents with one call, simplifying your search. if you need to filter down, most RAG setups have a concept of "metadata" which you could use to filter. going back to my example, if you know the document is about ford you could have a "company: Ford" metadata, then apply that filter when you retrieve. i know i blabbed a lot but hopefully that helps clarify things! please lmk if you have any questions on anything i said, happy to clarify 🙂

try prompting https://aiworkflowgenerator.xyz/ with a specific example using GoHighLevel 🙂

if you want to get started quickly, use https://aiworkflowgenerator.xyz/ to make workflows for you 🙂

so there's a few things you can do: 1. Prompt Caching - most big providers offer this, for OpenAI it's automatically done under the hood, but anthropic you have to opt-in by adding a specific header. the tldr is that if your prompts are usually the same, they can cache the part of the prompt that is identical across requests, which saves you some $ on input tokens 2. choose your model wisely - this is often why i advise people to separate out logic that is simple from a large model (i.e. sonnet 4, o3-pro) to a smaller model (gpt-4.1-mini, claude haiku). smaller models can be very good if the task doesn't require much intelligence and you have a few (7-10) good examples that you can guide the model with in the prompt. but the tldr is try to find ways to let smaller models handle tasks, you can get a lot more bang for your buck that way. even for larger models choose wisely - for example with the updated pricing for o3, i'd prefer that for most of my workflows over claude 4 sonnet/opus. sonnet is great but they get you on the output tokens, and o3 has impressed me thus far so i'd try that before starting with a more expensive model. you could also even try starting with deepseek or gemini flash, free models that are very good also. just my 2 cents though! let me know if anything i mentioned doesn't make sense lol, happy to explain more :)

in case anyone wants to check it out for themselves: https://aiworkflowgenerator.xyz/

@Austin Kory can you give me a specific example of the admin stuff you're trying to automate?