Skip to content
NovaVMS Manual

LLM pricing

NovaVMS v5 ships one VLM provider: Gemini 3.1 Flash-Lite via Vertex AI. Vendor list pricing changes; this page links out to the authoritative Google source and documents the cost telemetry NovaVMS exposes.

Authoritative price source

Always check the vendor’s pricing page before budgeting — NovaVMS does not re-bill or cache vendor rates.

In-app cost observability

Gemini judge consumption is tracked at the org level in two dedicated tables (migration 000076_gemini_judge_budget.up.sql):

TableKey columnsPurpose
gemini_judge_budgetdaily_limit (default 1000), alert_threshold_pct (default 80)Per-org daily budget configuration (operator-tunable from /settings/ai-budget).
gemini_judge_usageusage_date, judge_calls_count, last_call_atPer-org daily usage counter, reset via cron at UTC midnight.

A third table, gemini_judge_budget_alerts_sent, tracks which orgs already received an over-budget alert that day so admins are not spammed.

Both budget and usage surface on /admin/ai under AI Runtime — Model health in the Gemini card (frontend/src/components/AIModelHealthCards.tsx). When judge_calls_count exceeds daily_limit, the card renders in a degraded badge state and surfaces a “budget exceeded” tooltip. Search, indexing, and deterministic alerts continue regardless.

Cost levers in v5

These knobs reduce Gemini cost per event without changing search or NBD behaviour:

  1. Reduce vlm_understanding_enabled at the org level — skip the per-event summary entirely (events still index, search, and alert).
  2. Drop contextual judge usage by tightening compiled-rule prefilters so fewer rules require a Gemini judge call.
  3. Pin thinking_level_default = minimal at the org level — minimal is cheaper than low. Per-rule contextual judges can override.

The three CPU models (MobileCLIP2-B, YOLO26n, InsightFace) cost only host CPU; they do not bill per call.

Anthropic / OpenAI / Qwen note

Anthropic Claude, OpenAI GPT-4, and Qwen-VL were dropped as configurable providers in the L12 cutover. Adding any of them back would require new code in cloud/internal/ai/ plus a product decision.