Bounded clinical NLI batch planning¶

openmed.clinical.nli_batch_plan builds deterministic local-inference batches without inspecting premise or hypothesis text. Callers provide an opaque pair identifier and a token estimate produced by their selected local tokenizer. The planner performs no model loading, filesystem access, telemetry, or network request.

from openmed.clinical.nli_batch_plan import (
    NliPairCost,
    NliRuntimeProfile,
    plan_nli_batches,
)

profile = NliRuntimeProfile(
    max_tokens_per_batch=512,
    max_pairs_per_batch=8,
    max_total_tokens=1_024,
)
plan = plan_nli_batches(
    [
        NliPairCost("pair-001", 180),
        NliPairCost("pair-002", 420),
        NliPairCost("pair-003", 700),
    ],
    profile,
)

The planner preserves caller order and starts a new batch before either the per-batch token ceiling or pair ceiling would be exceeded. A pair larger than the per-batch token ceiling is returned in deferred_pair_ids. Optional total token and pair budgets are hard invocation limits; pairs that do not fit are also deferred while later smaller pairs remain eligible.

Pair identifiers are returned because the caller must be able to retry or route deferred work. They must therefore be opaque, non-PHI identifiers. The ordinary representations and to_audit_dict() omit identifiers; audit output contains only counts, token estimates, and domain-separated SHA-256 fingerprints. Exceptions never echo caller values.

Planning is a resource-control mechanism, not a clinical inference or safety decision. Deferred pairs require an explicit caller policy such as a smaller local profile, a later local run, or qualified review.