Context Packing

An agent's context window is finite. When it asks for context around a node, CGraph has to choose which of the surrounding graph to include — enough to be useful, few enough to fit. That is gather-and-pack.

Gather: fixed vs adaptive

fixed:     [ whole k-hop neighborhood ]        (query-independent)
 
adaptive:  [ full 2-hop core ]
              + third hop, only along query-relevant nodes

• gather: "fixed" packs the whole k-hop neighborhood — simple and query-independent.
• gather: "adaptive" keeps the full 2-hop core but expands the third hop only along nodes relevant to the query. It needs a query/q — without one the relevance gate is a no-op and it behaves like a fixed gather.

Pack: knapsack

The gathered candidates are packed to a token budget with a knapsack strategy (packing: "knapsack") — fit the most relevant nodes into the budget rather than truncating arbitrarily.

cgraph-client context '{"q":"Parser","budget":5000}'

Reach: an inspectable result

The response reports how the neighborhood was selected, so retrieval is auditable rather than a black box:

Alongside these, the response carries the gather and packing mode it used.

Why adaptive

Expanding every third hop is expensive; expanding none loses relevant context. Adaptive gather targets the middle: it spends tokens only where the query points. The measured tradeoff — recall lift versus token cost — is documented on the Benchmarks page with methodology, not asserted here.