Context Packing

Context packing is the process of selecting, ordering, and formatting information for a model call. It decides what evidence and state enter the context window.

Given candidate context items $c_i$ with utility $u_i$ and token cost $l_i$, context packing resembles a constrained selection problem:

$$\underset{S}{\max }\sum _{i\in S}{u}_i\text{subject to}\sum _{i\in S}{l}_i\le B$$

where $B$ is the available token budget after reserving instructions and output.

Context Item Types

• User request and current goal.
• System or repository rules.
• Current file excerpts.
• Retrieved documents.
• Tool results.
• Prior decisions and summaries.
• Output schema or examples.

Ordering Principles

• Put durable rules before task data.
• Put fresh task evidence near the action it supports.
• Separate untrusted content from instructions.
• Label sources and timestamps when relevance depends on freshness.
• Prefer concise excerpts over full dumps.

Checks

• Does each context item support the next decision?
• Are source boundaries clear?
• Is any private or irrelevant information included?
• Is the model being asked to infer from missing evidence?
• Would a smaller context produce a more reliable answer?

Related

• Token budget
• Context window
• Retrieval-augmented generation
• Chunking
• Query rewriting
• Context engineering
• Memory boundary