Ranking local services under live constraints

This worked application uses explicit planning assumptions to demonstrate FIELD. Quantities below are inputs, calculated implications or proposed targets; they are not reported customer results. The engagement decision remains conditional on the evidence described.

Planning basis and engagement boundary

F — Frame the outcome

Improve successful local transactions while respecting geography, availability and merchant participation at the stated serving scale.

Value measure: Successful eligible transactions, cancellations, user effort and merchant coverage.

I — Investigate the workflow

Observe failed searches, stale inventory and cold-start merchants. Establish business guardrails beyond clicks and model how marketplace participants affect each other.

First evidence task: Replay queries with unavailable high-score items and new merchants.

E — Engineer the solution

Use candidate/ranking/feature stages, freshness constraints and the million-scores-per-second capacity calculation; tie each stage to missed-result attribution.

The implementation assessment below makes the authority, workload, resource and failure assumptions concrete. Its numbers are calculations from the stated scenario, not a vendor performance guarantee.

Contract

Define eligible geography, availability, user intent and product outcomes: successful transactions, satisfaction/cancellations and merchant coverage, not clicks alone. Specify p99 deadline, inventory freshness and fallback if features are stale.

Implementation plan

Stream item/POI/availability changes into authoritative records and feature pipelines. Candidate retrieval combines geo/eligibility filtering with lexical/embedding sources. A lightweight scorer narrows candidates; a richer ranker scores 100/query and a policy layer enforces constraints/diversity. Log exposures and outcomes with model, feature and candidate versions.

Sizing

10M×768×4 gives 30.72GB raw vectors. At 10,000 queries/s×100 reranked candidates, demand is 1M candidate scores/s. Five uncached separate feature fetches/candidate could create 5M lookups/s, showing why batched/prejoined features and local caches matter. A measured ranker capacity of c candidates/s/device implies at least 1M/c devices before safe utilization/failure reserve; c cannot be invented from model size.

Learning

Use point-in-time labels/features and distinguish nonexposure from negative preference. Mix validated random/hard negatives and evaluate cold-start regions/items. Geographic eligibility and merchant state remain enforceable independently of scores. Content/metadata priors and bounded exploration support new merchants.

Failure scenarios

Stale availability triggers revalidation or exclusion under a freshness policy. Popularity concentration requires exposure/diversity analysis rather than blindly suppressing all popular items. Sparse regions may use broader retrieval or simpler priors with explicit uncertainty.

Evaluation and rollout

Measure candidate recall, ranking quality, serving tails and freshness first. Online tests include conversion, cancellation, user experience and merchant distribution, using a design appropriate to interference. Preserve fallback and a compatible model/feature bundle. A relevance gain without timely eligible inventory may reduce real conversion.

L — Launch with evidence

Run suitable offline and marketplace-aware online evaluation, including latency, cancellations and merchant coverage. Avoid optimizing click-through alone.

Critical decision condition: Ineligible/stale inventory or serving tails beyond the agreed contract block expansion.

Attach raw artifacts, the tested configuration, population, denominator, date and limitations to the release record. Missing or stale evidence holds the dependent decision. A favorable value forecast cannot offset a failed critical condition.

D — Develop ownership and reuse

Search product, inventory and ML platform owners share explicit freshness/model contracts; maintain fallbacks and experiment ownership.

Candidate reusable asset: Eligibility/feature contract and ranking-stage evaluator.

Commercial decision and next commitment

Accountable decision owner: Marketplace search outcome owner.

The initial commitment is the bounded journey above, with the stated exclusions. Confirm access, customer staffing and acceptance ownership during discovery; price and schedule unresolved dependencies explicitly. A new source, tenant class, language, action or operating envelope requires an impact review rather than silent scope growth.

Approve only the next exposure level supported by the evidence. The next useful investment is: replay queries with unavailable high-score items and new merchants.

Method and references

Apply the FIELD method, economic model and evidence rules. Methodological and technical influences are listed in the source register.