Sparkles

Appearance

GPU retrieval — inverted indexes and learned sparse models on device ​

The other half of the accelerator question: not "match this pattern faster" but "traverse these posting lists faster". Out of scope for the same reason ranked retrieval is, and recorded so the boundary is explicit.

Last reviewed: August 28, 2026.

NOTE

A boundary page, [literature] throughout. Nothing in this catalog's decision space depends on it.

What the work is ​

Two related lines:

Inverted-index traversal on GPU. Posting-list intersection and top-k selection map onto data-parallel hardware reasonably well — the lists are sorted integer sequences and the reduction is a bounded heap. The literature reports throughput gains for high-QPS serving.

Learned sparse retrieval. Models like SPLADE produce weighted term expansions, turning one query into hundreds of weighted terms. That inflates traversal cost enough that GPU execution becomes attractive, and it is where most current attention sits.

Why neither applies here ​

  1. The workload is queries-per-second, not latency-for-one. These systems amortise device occupancy across concurrent queries. A single-user editor issues one query at a time; there is nothing to batch.
  2. It presupposes an inverted index, which theory/ranked-retrieval.md concludes is the wrong model for code search, and Lucene and tantivy elaborate.
  3. It presupposes ranking by term statistics, and code relevance is structural — definition versus mention, frecency, git status, path distance.
  4. The platform prerequisite is missing anyway.

The one transferable idea ​

Bounded top-k selection is the same problem regardless of hardware. The GPU literature's contribution — maintain a bounded heap, use score upper bounds to skip whole blocks — is the CPU technique from PISA and the WAND family, and it applies to sparkles:fuzzy's existing TopK without any of the surrounding apparatus. That is already recorded in theory/ranked-retrieval.md; this page adds only that GPU work has not changed the underlying idea.

What this catalog concluded ​

Out of scope, with no open question left behind.

Sources ​

[literature]: the GPU inverted-index and learned-sparse-retrieval literature. The relevant CPU-side techniques are surveyed in PISA and theory/ranked-retrieval.md.