Context Store & Context Assembly

Retrieval finds candidates. Context Assembly determines what may shape the decision.

Combine eligible documents, operational data, calculated facts, features, memory, policy, user intent, and tool observations into one versioned Context Package.

Scope eligibility, authority, purpose, residency
Sufficiency ready, enrich, verify, human review, block
Retrieved evidence does not inherit instruction authority.
Decision boundary

Assemble the decision boundary, not just a search result

Instructions and intent

Trusted guidance, user intent, schemas, and decision requirements define what the boundary is for.

Facts and evidence

Retrieved documents, structured query results, feature values, operational state, and derived facts enter with roles.

Policy and observations

Policy facts, tool observations, memory, and prior context can participate without being flattened into prompt text.

Sufficiency gate

Know when context is not ready for judgment

A completed retrieval is not necessarily a complete decision boundary. Missing, stale, denied, conflicted, incomplete, and out-of-budget context stay distinct.

READYENRICHVERIFYHUMAN_REVIEWABSTAIN_OR_BLOCK
Semantic roles

Keep instructions, evidence, and untrusted content in distinct roles

Authoritative fact
Retrieved evidence
Derived feature
User assertion
Advisory memory
Policy fact
Tool observation
Trusted instruction
Eligibility first

Apply eligibility before information can influence the decision

A document can be highly relevant and still be ineligible for this user, workload, tenant, resource, purpose, data classification, region, or moment in time.

Continuous Authority

Carry authority into Context Assembly

Eligibility can change after permission revision, resource change, reclassification, expiry, policy change, or purpose change.

Package contents

Combine structured and unstructured context in the same package

A supplier bank-account-change decision may need current operational truth, retrieved evidence, calculated features, knowledge relationships, policy facts, and tool observations.

Operational facts from authoritative sources
Retrieved letters, requests, and verification evidence
Calculated account-change and risk features
External validation tool observations
Transformation lineage

Treat transformations as application behavior

Chunking, OCR, extraction, redaction, summarization, normalization, aggregation, and feature calculation can change what the decision actually sees.

Freshness contract

Keep context fresh enough for the decision

Effective time, observed time, permitted staleness, expiry, point-in-time requirements, and revalidation keep context from outliving its business meaning.

Map the Context Package for one consequential AI decision.

Use an architecture review to identify context roles, eligibility, sufficiency, freshness, transformations, memory boundaries, and provenance for one application path.