Oracle’s Select AI stores NL2SQL corrections as a vector index—not a fine-tune

Oracle has exposed a concrete answer to one of production text-to-SQL’s hardest questions: how does a system learn that finance means REV_AMT_NET, or that an “active customer” follows a company-specific rule, without retraining the model? In a new Autonomous AI Database on Dedicated Infrastructure walkthrough, Oracle says Select AI Feedback now lets operators attach positive or negative feedback—and a replacement SQL response—to an AI profile.

The correction becomes retrievable data

On first use, Select AI creates a vector index named after the profile, with a backing table that stores embeddings of prompts and corrections. When a later prompt arrives, vector search retrieves the closest examples and injects them as hints. Oracle says the default retrieval limit is three matches.

That makes the feature persistent retrieval-based prompt augmentation, not fine-tuning. No model weights change, and dropping the AI profile also drops its feedback index. The scope is deliberately narrower than a general memory system: Oracle documents feedback for the NL2SQL actions runsql, showsql, and explainsql, on profiles configured for SQL generation rather than RAG.

The distinction matters operationally. A correction can follow a changing schema because it remains a database object rather than becoming behavior baked into a model checkpoint. It is also inspectable and lifecycle-bound to the profile. But retrieval introduces its own test surface: teams should verify whether the right correction enters the top three, whether an obsolete correction is still retrieved, and whether similar-looking questions require different business definitions.

The profile is now the control point

The same release expands that profile-centered design beyond feedback. Oracle says Select AI can now use Amazon Bedrock models as SQL generators; AWS profiles must name a model or inference profile explicitly, and ADB-D operators remain responsible for the regional endpoint, network access-control entry and an actual outbound route. Oracle warns that an ACL grants permission but does not create a route from a private subnet.

Select AI also supports persistent, resumable conversations whose prompts are queryable through a database view, plus an in-database agent framework with built-in NL2SQL and RAG tools and custom PL/SQL procedures as actions. Those additions put corrections, conversation state and tool execution beside the data and its existing privilege model rather than in a separate application cache.

For ADB-D deployments, Oracle lists hard prerequisites and tradeoffs: the agent package requires Oracle Database 19c 19.29 or later, or 26ai 23.26 or later; application schemas need explicit package grants and per-principal network permissions; and in-database embeddings and parallel synthetic-data generation consume provisioned ECPUs.

The practical takeaway is not that feedback eliminates evaluation. It makes feedback testable. Before switching it on broadly, teams can build a regression set that records which correction should be retrieved, retire corrections when schema semantics change, and compare generated SQL with and without the feedback index. Oracle’s implementation gives that loop a durable home; production teams still have to govern what the system is allowed to remember.