PATTERN Cited by 1 source

Deterministic filter before LLM reasoning¶

Pattern¶

Place a layer of programmatic, rules-based checks before an LLM inference step. Items that match known-outcome patterns (benign, invalid, duplicate) receive an instant disposition without consuming model tokens. Only genuinely ambiguous items proceed to the expensive reasoning step.

When to use¶

High-volume pipelines where a significant fraction of inputs have predictable outcomes.
Cost-sensitive deployments where LLM inference is the dominant expense.
Latency-sensitive paths where sub-second response is needed for most items.

Structure¶

Input stream
    │
    ▼
┌─────────────────────┐
│ Deterministic rules │  ← trusted lists, regex, lookups
│ (30-95% handled)    │
└──────────┬──────────┘
           │ unmatched (ambiguous)
           ▼
┌─────────────────────┐
│ LLM reasoning       │  ← expensive, slow, probabilistic
│ (5-70% of volume)   │
└─────────────────────┘

Key insight¶

"The cheapest call is the one you never make." At Databricks, deterministic filtering handles 30–95% of security alert volume per source, saving thousands of LLM calls per day (Source: sources/2026-07-06-databricks-scaling-security-alert-triage).

Seen in¶

sources/2026-07-06-databricks-scaling-security-alert-triage — Databricks security triage pipeline: deterministic filters before 17 source-specific LLM agents.