0025 — Gemini 3.1 Flash-Lite, called on the global endpoint¶

Status: accepted (for V2; supersedes 0009)

ADR-0009 chose Gemini 2.5 Flash-Lite to avoid a retiring family. The 2.5 family now retires in turn (shutdown 2026-10-20), so its own argument applies to it. The reasoning is unchanged — extraction and scoring are simple structured tasks, so the cheapest capable tier wins — and only the id moves: gemini-3.1-flash-lite.

The model has no regional endpoint. One generateContent call per endpoint, 2026-08-15:

Endpoint	Result
`global`	200, `modelVersion: gemini-3.1-flash-lite`
`northamerica-northeast2`	400 `FAILED_PRECONDITION`
`northamerica-northeast1`	404 `NOT_FOUND`
`us-central1`	404 `NOT_FOUND`

No foundational Gemini model runs natively in northamerica-northeast2, the original warehouse location, and Google's in-region Canadian processing (northamerica-northeast1) does not serve this model either — so the choice was between the global endpoint and no model at all. Global is acceptable: the pipeline reads public job postings and has no residency requirement.

BigQuery's supported endpoints are not Vertex's, and that is the practical constraint. us-central1 returns 404 for this model at the raw publisher endpoint, yet AI.GENERATE reaches it from a us-central1 dataset — BigQuery resolves through its own path against its own allowlist. Probed by calling AI.GENERATE with each endpoint, 2026-08-16:

Endpoint	BigQuery
`gemini-3.1-flash-lite`	works — the newest flash-lite it accepts
`gemini-3.5-flash`	works (newer, but the flash tier)
`gemini-3.5-flash-lite`	`Unsupported endpoint`
`gemini-3.1-flash`	`Unsupported endpoint`

The allowlist is non-contiguous, so "newest available on Vertex" is the wrong question. Probe BigQuery directly before changing scoring_endpoint, and prefer an explicit version over the gemini-flash-lite-latest alias, which resolves but changes the model under a column of scores that claims to be comparable.

First backfill, measured 2026-08-16 (1,742 distinct posting texts):


Extraction	1,742/1,742 succeeded, 200s
Scoring	1,742 rows, 203s
`requirement_text`	570 chars average, against 6,854 for the full posting — a 92% trim
Second incremental run	0 rows, confirming the `content_hash` guard
BigQuery compute, all AI jobs	0.131 GiB billed, 100.8 slot-minutes — under a cent
Out-of-range generations	2 in 1,742 scored outside 1–5

The dollar cost is not measured and cannot be from here. Gemini token charges bill through Vertex and never appear in INFORMATION_SCHEMA.JOBS — only the BigQuery compute above does — so the authoritative figure is the billing console, the same conclusion ingestion-cost reached about storage rates. Measured input volume is ~3.0M tokens per extraction pass and ~3.1M per scoring pass, which bounds it at sub-dollar even at triple 0009's rate.