New 26ai features, tested on Oracle Database Free, with the catch
Every Oracle release ships a long list of new features, and the list is the easy part. What matters before you rely on one is how it behaves on a real database: what it needs, what it changes for the sessions already running, and where it stops short of the announcement. Each post in this series takes one 26ai feature, runs it on Oracle Database Free in a Docker lab you can rerun, and reports what actually happened, including the catch.
Lab On Oracle AI Database 26ai Free as shipped, the HNSW index failed with ORA-51962 and the vector pool couldn't be raised without a restart. Sized at the documented formula it still failed. After a restart with a pool smaller than the index, the index stayed VALID and every vector query quietly became a full scan.