AI functions reach the SQL standard

At the October meeting of the SQL Standards committee on October 13, two AI topics were on the agenda. I attended the meeting as a guest presenter and it was a great experience. EDB colleague Peter Eisentraut and I presented a discussion paper on model inference functions in SQL. 

The paper we presented (PDF version / Markdown) proposes how to specify a library of these functions: a structure for the library, a fixed template for each function, and shared concepts such as model references, error conditions, and determinism. One example: today, a query can compare vectors from two different embedding models and return a meaningless ranking with no warning. The paper requires an error in this case. 

The reaction from the committee was very positive. Several members suggested the paper belongs in ISO/IEC 29075, the SQL function library standard, and encouraged us to continue. 

Another working group participant presented a proposal, which adds an AI_GENERATE function to SQL. It sends a prompt to a registered LLM and returns the result. The committee accepted it for SQL/Foundation, the core SQL standard. 

AI functions are now becoming part of standard SQL. The two papers fit together. AI_GENERATE gives SQL one standard AI function. Our paper describes how to specify the rest: embedding, aggregation, and ranking functions. So: more to come! 

Peter occasionally posts about his work on the ISO/IEC JTC1 SC32 WG3 ISO/IEC JTC1 SC32 WG3 “Database Languages” committee, to other people just known as the SQL Standards Committee, on his blog. He joined the committee back in 2019. For details on the committee's operations, meeting cadence, and opportunities to get involved, check out Peter’s PGConf EU 2023 presentation. 

It’s not the first time EDB colleagues have joined the working group. In March this year, Matheus Alcantara and Euler Taveira joined the meeting in São Paulo, Brazil.