FAQ

Is this a prompted persona on a general LLM?

No. EGGai‑v1 answers as a person, grounded in the real till history of a major SEA grocery retailer, and every reply is validated against what those shoppers really did next — so it can be marked right or wrong, which a persona prompt on a general LLM cannot be.

How do I know the answers are any good?

Each SDK rests on a fielded study with a held-out topline the model had never seen, written up in full on the research pages — including where a purpose-built system beats us. The headline checks: 82.0% accuracy on stated willingness to pay (frontier model 67.1%); a redeemer ordered above a non-redeemer 76 times in 100 across 12,480 held-out offers; 64% of panel answers landing where 1,013 real respondents landed theirs (frontier AI 52%).

Why does the survey panel disagree with itself?

By design. The panel is built to match the distribution of real answers, never to sharpen onto one confident answer. A panel that agrees with itself more than real people do is wrong about people in a way accuracy alone will not catch. Read the spread — and field a real panel on the questions that still look uncertain.

Why did my offer optimization run return nothing?

The guidelines ran out before the judge accepted. That is a verdict, not an error: inside your caps, there was no version of the offer worth making for this customer. Loosen the guidelines only if the economics genuinely allow it.

Can the optimizer invent an offer that breaks my rules?

No — twice over. The proposer opens at the stingiest version your guidelines allow and never goes outside them; prohibitions you state ("never promise free delivery") bind the wording as well as the terms. And numeric constraints are enforced by the harness itself, so terms outside them never come back regardless of what any model writes.

How is an assortment answer different from a demand model?

A demand model tells you what the lines sold; this tells you what the people do next — one outcome per customer, with the spend attached and the walkers named. Prices hold and nothing is on promotion for the run, so the shelf change is the only thing that moved. And it is calibration, never causal proof: sealed rows are scored against your actuals for closeness, not credited with causing them.

What data do I need to provide?

Offer Optimization and Lead Generation run on customer profiles — best generated from transaction history onboarded into your workspace, so each customer_id resolves to a twin. Assortment Optimization runs on a panel of the category's customers, ideally with 28-day spend columns. Survey Response needs nothing — the panel is drawn from real purchase histories for you.

Do twins see each other's answers?

No. In rankings, panels and assortment runs, each twin answers independently. A ranking is many independent judgements, and a panel's spread is 150 separate people, not one model sampled 150 times.

Why is Survey Response marked "preview"?

The survey endpoints are being rolled out per workspace. The SDKs ship the full interface now; where your workspace is not yet enabled, calls raise FeatureNotEnabledError. Ask your EGGai contact to switch it on.

In what language does the model reason?

In the shopper's own language and register, translated for display where needed. The model is calibrated to how people in this region actually answer.

Can I use it outside grocery / outside SEA?

EGGai‑v1 is a regional grocery behaviour model; that grounding is where the accuracy numbers come from. Talk to the EGGai team about your domain before assuming transfer.

Where do I get an API key?

In the EGGai app, under Settings → API keys — one live key and one practice key per workspace. Keys are scoped to the workspace, which is also where your data, jobs and metering live. No app access yet? Your EGGai account manager can set the workspace up.

Is there a sandbox?

The practice key is the sandbox: it authenticates everywhere the live key does, runs the same models, and caps rows per job so integration mistakes cost nothing.