Expand description
Explicit graph-assisted inference ABI.
Graph retrieval is a cold/request-bound workload reducer, never an implicit token-kernel side effect. Results, token spans, and prefix-page identities are caller-buffered and separately accounted so model throughput cannot be relabelled as retrieval throughput.
Structs§
- Graph
Assist Policy - Graph
Query - Graph
Selection Receipt - Prefix
Identity - Prefix
Page Registry - Fixed-capacity prefix-page index. Lookup/insert/evict never allocate.
- Prefix
Page Set
Enums§
Functions§
- derive_
prefix_ identity - Bind reusable KV pages to the exact model, tokenizer, graph scope/revision and selected facts.
- query_
graph_ indexed_ into - Select graph facts using a pre-built
QuinIndexto scan only the narrowest non-wildcard posting list in ascending graph order ($O(k)$ candidate inspection instead of $O(N)$ full scan). - query_
graph_ into - Select graph facts directly from the flat Quin store into caller-owned storage.