Context Pipeline: A Two-Stage Framework for Enterprise Context Readiness and Pre-Inference Context Compilation

Enterprise applications of large language models (LLMs) depend not only on model capability but also on the quality, authority, and fitness of the information supplied at inference time. In enterprise analytics, a model may receive a user question and retrieved records while lacking essential qualifying knowledge: which source is authoritative, how a metric is defined, which relationships and table grain are valid, whether data is current, what quality limitations apply, and whether use is permitted. Conventional retrieval can improve topical relevance, but relevance alone does not establish that information is approved, current, valid, or authorized for a particular task. This paper proposes Context Pipeline, a conceptual two-stage framework for preparing enterprise context before LLM inference. The first stage, Enterprise Context Readiness, maintains a reusable knowledge foundation comprising technical metadata, business semantics, relationships and grain, data-quality status, freshness, provenance, and governance constraints. The second stage, Task-Aware Context Compilation, interprets a task and assembles a validated, prioritized, budget-aware Context Package from that foundation. The package explicitly distinguishes confirmed facts, approved definitions, observations, inferences, and unresolved issues. It is designed to preserve traceability and prevent uncertain information from being silently represented as authoritative. The paper positions the proposal in relation to context engineering, retrieval-augmented generation, context compression, agent memory, and enterprise text-to-SQL research. It does not claim that the individual capabilities are new. The candidate contribution is their integration into an enterprise-oriented preparation lifecycle, with quality, freshness, provenance, and governance treated as context eligibility conditions during compilation. The paper formulates research questions, testable hypotheses, and an evaluation design comparing raw-context, conventional retrieval-augmented, and Context Pipeline conditions. The framework has not yet been implemented or empirically evaluated; its novelty and effectiveness remain to be established through systematic review and experimentation.