Skip to the document
Madhuopen lab

AI Insights: Latency & Cost Optimization · part 16 of 17

needs more data3 August 2026cost: API-only

Hybrid mode: short facts summary vs the raw ORM

Can the model add one grounded observation the rules missed — and what data should it see?

The question

Rules can't anticipate every pattern. Can the model contribute one extra 'AIObservation' — and does it need the raw data or just a summary of it?

needs more data

A short 3-year facts summary works: the model found the income trend in 2/2 runs. Feeding it the full raw ORM found nothing in 2/2 runs at 5x the tokens. Hybrid's value on thin returns still needs data.

Summary observation rate2 / 2 runsflagged 80k→100k income trend
Raw ORM observation rate0 / 2 runsmostly-null JSON hides the signal
Summary cost+865 tok~$0.0092/call total
Raw ORM cost5,228 tok5x tokens, worse output

How hybrid works

The polish prompt additionally carries a short summary of the customer's numbers (the SignalDigest in the code) — a 3-year view of all metrics plus the W-4 profile, +865 input tokens. The model may append at most one 'AIObservation' insight, with numbers quoted verbatim from that summary, confidence capped at 75, consultative tone. It worked: the model flagged a 3-year income trend ($80k → $100k) that no threshold rule covered.

ORM-as-input: rejected

The obvious alternative — just send the full raw ORM and let the model find patterns — failed cleanly. The full ORM costs 5,228 prompt tokens and the model added no observation in 2 of 2 runs: a mostly-null 59 KB JSON hides the signal. The short summary surfaced a trend in 2 of 2 runs at a sixth of the tokens. Higher cost, worse output — don't send the ORM. (The --orm mode is kept in call-ai.js for re-verification.)

Cost comparison

Cost comparison
ModePrompt / completion tokensLatencyCost per call
Regular (rules only)1,034 / 495~5 s~$0.0062
Hybrid (+summary)1,891 / 685~6 s~$0.0092
ORM-as-input5,228 / —rejected: no observations produced
hybridSignalDigestcontext designORM