Release · 2026-08-14T22:37:15+02:00 · By Threadnote

Threadnote 4.2.5

Threadnote 4.2.5 makes vector-enabled code graph builds faster on multicore CPUs while keeping GPU workloads and interactive semantic queries conservative by default.

What's new

Threadnote 4.2.5 makes vector-enabled code graph builds faster on multicore CPUs while keeping GPU workloads and interactive semantic queries conservative by default.

Faster CPU graph embeddings

  • Code graph vector materialization now shares one loaded model across up to eight independent embedding contexts and divides the complete detected math-core thread budget between them. Smaller CPUs automatically select the largest supported pool that fits.
  • On the governed eight-math-core, 10,000-symbol benchmark, the upper-median cold index time fell from 61.84 seconds to 42.40 seconds, and the vector stage fell from 50.72 seconds to 30.99 seconds. Eight contexts won all four paired rounds with a 1.49× upper-median paired-run speedup; ordered symbol-to-vector digests stayed identical, and upper-median sampled embedding process-tree RSS increased by about 6%.
  • Vector ordering, incremental-versus-fresh-rebuild parity, and the previous active mapping remain protected when a build fails or is interrupted; a retry can safely rebuild the pooled session.

Conservative accelerator and query behavior

  • Models with GPU layers, or an unknown offload state, continue to use one embedding context. Recall and semantic code graph queries also remain serial by default, so the graph-build optimization does not add unmeasured cold-query or accelerator contention.
  • THREADNOTE_EMBEDDING_CONTEXTS=1, 2, 4, or 8 remains a process-wide override. Prefix only the graph command to scope a low-memory rollback to that build:
THREADNOTE_EMBEDDING_CONTEXTS=1 threadnote graph index

Existing standalone installations upgrade with:

threadnote update