@memberjunction/feature-pipelines-engine

Server-side execution engine for Knowledge Hub Feature Pipelines (WorkType = 'Infer' on MJ: Record Processes). All Feature Pipeline runtime logic — inference drivers, MJ: Feature Values persistence, version activation & backfill stewardship, and golden-set test evaluation — lives here, decoupled from the generic Record Set Processing, Scheduling, and Testing Framework substrates.

Pure specification types, constraint/escalation validators, pin hashing, and FeaturePipelineRunResults live in @memberjunction/feature-pipelines.

Architecture & Substrate Seams

Surface	Registration / Entry Point	Description
WorkType = 'Infer'	FeaturePipelineStartup (@RegisterForStartup) -> RegisterInferRecordProcessor()	Registers the 'Infer' factory in RecordProcessorRegistry.Instance on module load and startup
Pipeline Drivers	LLMFeaturePipelineDriver, DecisionFeaturePipelineDriver (BaseFeaturePipelineDriver)	Turn a record’s context into structured output values via AIPromptRunner or AIDecisionRunner, with optional borderline escalation
Record Processor	InferProcessor (FailRunOnRecordError = true, BuildRunSummary)	Enforces version activation gates, resolves lookup taxonomies, validates output constraints & confidence floors, upserts MJ: Feature Values via FeatureValueStore, and delegates write-back
Version Steward	FeaturePipelineVersionSteward + StewardFeaturePipelineVersionAction (@RegisterClass(BaseAction, 'Steward Feature Pipeline Version'))	Evaluates unevaluated gated versions and runs the one backfill of an active changed version; invoked per pipeline by the scheduled 'Feature Pipeline Version Steward' MJ: Record Processes row (WorkType = 'Action', every 15 minutes via RecordProcessScheduledJobDriver)
Golden Set Driver	FeaturePipelineGoldenSetDriver (@RegisterClass(BaseTestDriver, 'FeaturePipelineGoldenSetDriver')), GoldenSetScorer, GoldenLabelQualityCheck	Evaluates a Feature Pipeline Golden Set test against the pipeline’s current version in one dry-run batch (BypassCache: true) and records the evaluation through FeaturePipelineVersionService

Feature Pipeline Results (ResultSummary)

InferProcessor upserts each record’s Feature Values on their key (ValueKeyHash). Its RecordResult carries Outcome (Unchanged or Upserted), PipelineVersion, HumanValuesPreserved, RejectedInvalid when an output violates its constraint, and Warnings (stored in MJ: Process Run Details.Warnings).

At the end of the run, InferProcessor.BuildRunSummary builds a FeaturePipelineRunResult (via FeaturePipelineRunResults.Build in @memberjunction/feature-pipelines) and stores it in MJ: Process Runs.ResultSummary:

• The pin (PromptVersionHash, PrimaryModelID, ParametersHash, SpecHash) and PipelineVersion the run used
• Tallies: Attempted, Upserted, Unchanged, Previewed, RejectedInvalid, BelowConfidence, Failed, Skipped, HumanValuesPreserved
• Because InferProcessor.FailRunOnRecordError = true, a Feature Pipeline run with any failed or rejected record ends Failed.

Dry Runs

An Infer dry run writes nothing (no Dedup Cache entry, Feature Value, watermark, target or child row) but still calls the model, so its MJ: AI Prompt Runs are recorded and cost what a real run of those records would.

• Without a scope or a single record it previews at most 25 records (DryRunSampleSize overrides it).
• A record whose real write would fail is reported Failed with the write-back’s reason (PreviewErrors).
• It reports, in WouldRefuse, why a real run would be refused (a version that has not passed its golden set), and never registers or activates a version. RecordProcess.RunNow / PredictiveStudio.RunFeaturePipeline return it as wouldRefuse; the Bulk Operations dialog then offers no Apply.
• Each record’s previewFields / previewChildren / previewTags hold only what the real write would change. Outputs withheld below their confidence floor are listed in withheldFields (and withheldKeptFields when the column does not allow null) and withheldChildEntities.
• An output with no Target (“Feature value only”) writes nothing to the record; its dry-run result is previewFeatureValues (FeatureName, Previous, Proposed), listing the Feature Values a real run would store.

Golden-Set Evaluation & Label Quality

• FeaturePipelineGoldenSetDriver — Driver of the Feature Pipeline Golden Set test type. Verifies that the pipeline names the test as its Activation.GoldenSetTestID and that model-made labels pass GoldenLabelQualityCheck, runs the current version as one dry-run batch with the cache bypassed (AI Prompt Runs are recorded), scores the answers with GoldenSetScorer, and records the result on that version through FeaturePipelineVersionService, which activates or refuses it against the pipeline’s Activation threshold.
• GoldenSetScorer — Pure scoring of a golden set: per-output accuracy with alternates accepted case-insensitively and a Wilson 95% confidence interval, abstention cases (ExpectAbstain), and overall pipeline value equal to its lowest output accuracy (or 0 when any Required case is wrong).
• GoldenLabelQualityCheck — Automated quality check for model-made golden labels: Adjudicate builds cases from labelers A, B, and C (A–B agreement keeps A, C adjudicates disagreements, three-way disagreements are excluded) and Check verifies a stored InputDefinition.LabelQuality record (two independent vendors, pairwise Wilson 95% lower bound >= 0.70, <= 5 of 52 excluded, and SHA-256 of cases) before any model call.

Version Stewardship

• FeaturePipelineVersionSteward — Evaluates a gated pipeline’s unevaluated version (running its Activation.GoldenSetTestID) and executes the one backfill of an active changed version (records whose latest value is under another version or none, with SkipUnchanged: false so watermarks are refreshed). A pipeline can opt out per step via Unattended: { AutoEvaluate: false } / { AutoBackfill: false }.
• StewardFeaturePipelineVersionAction — MJ Action (DriverClass = 'Steward Feature Pipeline Version') that runs FeaturePipelineVersionSteward.Instance.StewardPipelineByID(recordProcessID, { ContextUser, Provider }) for a single pipeline and outputs Outcomes, Version, LastAction, and Detail.
• Scheduled execution — Seeded as the 'Feature Pipeline Version Steward' MJ: Record Processes row (WorkType = 'Action', Entity = 'MJ: Record Processes', ScopeFilter = "WorkType='Infer' AND Status='Active'"), dispatched every 15 minutes by the 'Feature Pipeline Version Steward' MJ: Scheduled Jobs row (RecordProcessScheduledJobDriver).