experimental

Experimental (NOT deployed)

These statements explore an agentic remediation branch — using Confluent's built-in AI_COMPLETE (Gemini, via CREATE CONNECTION + CREATE MODEL) to turn efficiency events into natural-language remediation recommendations.

They are kept here for reference but are not part of the deployed pipeline, and the main README does not claim them.

Why they're not deployed

AI_COMPLETE is a non-deterministic function. Confluent Cloud for Apache Flink only allows a non-deterministic function over an append-only (insert-only) stream — it rejects it over the changelog (retract/upsert) streams that the windowed ML functions and dedup queries produce ("...can not satisfy the determinism requirement for correctly processing update messages").

An append-only window aggregation (GROUP BY window_start, window_end) feeding AI_COMPLETE into an explicit no-primary-key sink does reach RUNNING, but in our testing did not reliably materialize output rows within the demo window (async model-inference + windowing). Rather than ship a node that isn't demonstrably producing data, we left it out (no red/empty nodes in the lineage).

A production version would likely use Confluent Streaming Agents (AI_RUN_AGENT / CREATE AGENT), which are designed for this, or stage events to an explicitly append-only topic first.

Files

• 06a_gemini_connection.sql — CREATE CONNECTION to Google AI (Gemini). The API key is a placeholder; in practice the connection is created via the Confluent CLI so the key is never written to a file.
• 06b_gemini_model.sql — CREATE MODEL (text generation) for remediation.
• 06c_remediation.sql — AI_COMPLETE over an events stream.
• 08_events.sql — a unified/idle event stream intended to feed the remediation.

No API keys or secrets are stored here.