Google Spreadsheet + GPT3

And it’s even deeper. You’re touching on the DevOps side of employing generative AI in actual business cases. But not DevOps in the traditional sense; this is training curation DevOps. :wink:

The advent of usable AI models by domain experts has made obvious the need for operationally efficient patterns to help everyday domain experts and business people organize, curate, and transform their information into something AI models can digest. From Google sheets to Airtable, email messages, and PDFs - all of it will vie for inclusion into the future training sets we craft to put these new AI tools to good use.

Consider the analytics you need to know if your corpus curation process is improving the experience. This is exactly the approach I used with Coda (instead of Google Sheets) to craft something useful for my team.

Your suggestion of a Chat Completions sheet and versioning attributes is critical. My Coda-based process factors all these variables into the workflow.

But the workflow also needs to support a way to test quickly and with an equal measure of refinement. Model curators should be able to quickly assess and change the corpus and then try to ensure there’s no regression.

Google Sheets is a good place to start, but my experience has shown it is not likely to carry us to the finish line. You need an environment that support content development with deep tentacles into other data sources. It needs to be as intelligent as the AI solutions you are trying to build. This has caused me to see a clear indication that GPT itself, is needed in many aspects of the corpus creation process.

Example, when I discover a poorly performing query for an FAQ system, I need to assimilate perhaps two other vectors that are performing well and which each contain parts of the correct answer. Using Coda’s relationship capabilities and OpenAI completions, I can automate the assimilation of a new answer that addresses the poorly-performing vector. This is made possible by using keyword extractions that align the other answers and completions that paint the new answer. With a few final edits in place, I now have a new vector and all the data needed to know that this new thing needs to be tested.

1 Like