Selectively Grounded Generaton
Ozgur (Oscar) Ozkan
Software Engineer Contractor ( Bootstrapped an AI company to 1M$+ ARR in 3 months as the sole developer / Open to new opportunities ) SRE / Platform Engineering / DevOps / DevSecOps / Backend / Full stack
What's your no: 1 problem these days? I will explain:
Forming a good prompt every single time to communicate to ChatGPT in the right way is important. But there's an issue: Your prompts are stored in your garbage notes, and your data is in a folder named "unnamed folder." Well, good luck finding all that and constructing a prompt to get your answer.
Today, we are introducing you to Selective Grounded Generation. Well, what is it?
AI technologies surrounded by "Search" search for a needle in the haystack and use the needle to get quick grounded answers. Okay, this is cool, but it's level 0. What's next is the Selectively Grounded Generation. When synthesising new information, you often use knowledge from two different domains.
Take this example: you would like to get a profitability overview for your company, but you only have two different files:
File 1: Revenue
领英推荐
File 2: Costs
When you want to derive Profitability, which is Revenue—costs, you have to construct your prompt so that you point to specific parts (bookmark), copy and paste them to your inline chat, find costs from the costs file, copy and paste them to your inline chat, and then write a prompt.
You won't find any results when you search for "Profitability" unless it's directly stated in these documents. This is where you need to combine/merge two sources to come up with a prompt. You can use the revenue structure from file one and the cost structure from file two and ask ChatGPT about your company's profitability.
This is, ladies and gentlemen, what I call a selectively grounded generation. Your question is still grounded in two files, but their parts have different search queries. You have to find costs on the costs file, and you have to find revenue on the revenue file. So, AI Search, per se, is useless in these scenarios.
This is why we launched Keymate Memory PDF Drive; by allowing you to Search & Mix, and match different queries and files by selecting your pdf, you can ground ChatGPT answers using selectively grounded answers.
Try our new PDF Drive & Memory here: https://keymate.ai. It uses the latest OpenAI embedding model and works well.
Also, on another point, we made all usages on Keymate AI GPT on ChatGPT free and unlimited until the 1st of April; no, this is not an April Fools' Day joke.
Software Engineer at High Bee Solutions
8 个月Very nice Ozgur