I am no AI novice, with a 20-year background in setting up teams for 'generative' AI projects. So the current wave of transformer-based AI is both very exciting and strangely expected.
However, the current wave of generative AI from the likes of OpenAI are just scratching the surface. When these types of services are enriched with client-owned data, and search and processed along with the generative AI, you get MUCH better results and can perform deep investigations into large sets of information. This is known as RAG (retrieval-assisted generation). RAG is not too hard to implement... if you are a developer and understand how to refine how data gets 'embedded' into vector databases. And that is not most of us.
AfforAI looks to be a ready-made RAG service for people who do not want or need to know what is going on behind the scenes. To test AfforAI, I uploaded 300+ research papers (mostly PDFs) relating to a very tricky question about corporate different types of training programs and staff retention. Using AfforAI, I was able to confirm that while there was a strong correlation between training in general and staff retention, it was not possible to determine which types of training were most effective because there are too many other factors to consider. This was pretty much what I expected, but it was great to see how about 3 weeks of 'readings' could be condensed into 4 hours (most of which was sourcing research from journal databases).
So AfforAI is brilliant for research.
I then decided to try another scenario. I uploaded about 3mb of text relating to a TV show: characters, plot lines, etc. I then asked AfforAI about the show - in most (not all) cases the answers were spot on. Then I asked the AI to create new characters and episodes. It absolutely shone with this task. No. I am NOT saying it should write the scripts for the shows! This was a test, and a very successful one at that.
Finally, I started to use some of the prompts I use with OpenAI calls. The quality of response was on par with, and in a few cases better, than GPT 3.5.
Now that AfforAI allows you to add your own AI keys to the service, it should be possible to leverage other generative AI services in the future. I look forward to this.
My only wish is that the service would use storage limits rather than a limit of the number of documents uploaded. I have about 4,000 small research documents from my team I'd love to upload, but am limited to 1000 docs at a time.
I would also like to see a more robust file management capability. Loading documents into folders, and allowing folders to be deleted, added and removed from agents, etc. I am sure those more robust features will come. In addition, I'd like to have the ability to refine the document ingestion parameters, including the ability to set some meta data.