AI Context: Master the Information Environment Behind Generative AI

AI Context is the set of information made available to a generative model during a specific processing step and plays a central role in determining how a request is interpreted. In this handout, we will examine Context Engineering, which involves selecting, retrieving, organizing and updating the information environment surrounding the model.

We will then explore context management, including the distinction between persistent and local information, the need to preserve provenance and the risks created by overload, redundancy, conflict, obsolescence and contamination. Finally, we will analyze documentary context, which provides the materials and sources used for a task, and tool context, which describes the available actions, permissions, parameters and results that allow the system to interact with external resources.

Context

Fullscreen Download

Key points

Execution context = information actually made available to a language model during processing, including instructions, conversation history, documents, retrieved data, and tool results.

Internal knowledge ≠ context: internal knowledge consists of patterns learned during training, while context provides task-specific information during execution.

Available information ≠ execution context: resources stored by a system become context only when selected, retrieved, and presented to the model.

Prompt Engineering ≠ Context Engineering: the former designs the immediate request, while the latter manages the broader information environment supporting its execution.

Context quality ≠ context quantity: effective context prioritizes relevant, current, consistent, and clearly organized information rather than maximizing available content.

Persistent context ≠ local context: persistent context retains reusable preferences and conventions, while local context contains information specific to the current task or conversation.

Personal context = stable user preferences and characteristics that guide interaction across conversations without rigidly determining individual responses.

Conversational context = information, decisions, definitions, and conventions accumulated within a dialogue, enabling continuity while potentially preserving outdated assumptions.

Instructional context = active directives specifying objectives, operations, priorities, constraints, materials, output formats, and quality criteria for the current task.

Normative context = higher-priority constraints governing permissible system behavior, including safety, privacy, authorization, and data protection, regardless of user preferences.

Documentary context = documents, data, and other materials supplied or retrieved to support a task; their contents do not automatically constitute authorized instructions.

Tool context = descriptions, permissions, parameters, results, and errors associated with available tools; tool capabilities ≠ information describing those capabilities.

Context priority → personal → conversational → instructional → normative; this represents the handout’s functional model, not a universal technical instruction hierarchy.

Documentary and tool contexts → support task execution by providing information and operational capabilities without independently determining the system’s objectives.

Context lifecycle → selection → retrieval → organization → use → updating → compression or removal, ensuring information remains relevant throughout execution.

Information provenance = identification of where information originated, how it was derived, and whether it has been verified or remains uncertain.

Context contamination = process through which incorrect information or model-generated hypotheses become accepted as established premises in subsequent interactions.

Context degradation → may result from overload, redundancy, conflicting instructions, obsolete information, contamination, lost provenance, reduced focus, or poorly visible priorities.

Context maintenance → preserves valid decisions, highlights critical constraints, removes obsolete information, and distinguishes verified facts from provisional assumptions.

Context management ≠ output verification: context supplies the information needed for generation, while verification assesses its reliability and correct interpretation.