0 Releases 11 Tags

RSS Feed

• v0.8.0 37fdd33b2d

  Compare

  v0.8.0 — chunk-before-embed for arbitrarily long markdown (#38)

  mathias released this 2026-05-19 20:00:05 +00:00 | 22 commits to main since this release

  ChunkMarkdown splits at H1/H2 boundaries, sub-splits oversized sections
  at paragraph boundaries with greedy packing under maxChunkBytes=4000
  (≈1000 nomic tokens — well under the 2048 ceiling).

  Storage: each chunk lives at "#NNNN" in brain_embeddings, 1-based
  4-digit zero-padded for stable sort order. No schema change.

  Retrieval: hybridMerge collapses chunk-path vector hits to parent via
  ParentPath before scope check, RRF accumulation, and hydration. Three
  chunk hits → one result row.

  Backward compatibility: pre-existing bare-path rows in brain_embeddings
  keep working — ParentPath is a no-op for them. No migration needed.

  First production sync after deploy hit added=32 deleted=0 errors=0 —
  first errors=0 cycle in days. The three previously-failing files now
  have 9 / 11 / 12 chunks each, all retrievable via brain_query.

  Closes infra#38.

  Downloads

  • Source Code (ZIP)
  • Source Code (TAR.GZ)