Architecture

Developer source → extraction → heading-aware chunks → sharded BM25F index
                                          ├─ static vectors (CI) ─┐
                                          └─ runtime vectors ─────┤
                                      portable assets → browser ranking

The boundaries are deliberate:

• core knows the portable format and browser ranking, but not build tools.
• build owns the complete indexing pipeline.
• vite waits for generated HTML and invokes build; it has no indexing logic.
• the first-party Rust/Wasm embedding engine can run in Node and the browser.
• every corpus has its own metadata, postings, and either static, runtime, or no vectors.

This means a non-Seekite producer can emit the documented assets and use the runtime, while another runtime can consume assets produced by Seekite.

Format v2 eagerly loads a compact columnar corpus and front-coded dictionary, then fetches only the postings and content shards needed by a query. Static vectors default to per-row int8 quantization. Query analysis folds diacritics, optionally stems, expands prefixes and typos, then applies filters, document grouping, pagination, hydration, and highlight metadata.

Hybrid ranking normalizes per-corpus BM25F scores and combines them with cosine similarity. semanticWeight controls the blend and defaults to 0.5. The same runtime can live in a Web Worker through workerSearch, keeping inference and scoring off the main thread.