Architecture

Seek.js treats AI search as a static asset and edge compute problem—not a always-on vector database. The pipeline splits work across three places:

1. Build-time indexing

During your site build, @seekjs/parser extracts semantic text from rendered HTML (WASM-backed parsing), binds chunks to canonical URLs for citations, and streams batches to @seekjs/compiler. The compiler vectorizes chunks, validates schema, and emits a compressed binary index (.msp) you ship like any other static file.

2. Browser hybrid search

@seekjs/client (and framework bindings such as @seekjs/react) downloads the .msp from your CDN, caches it in IndexedDB, and runs hybrid search (BM25 + vector similarity) entirely in the browser—targeting sub-15ms query latency for local retrieval.

3. Edge reasoning (optional)

When a user wants a generative answer—not just search hits—@seekjs/ai-edge runs on Cloudflare Workers (or similar): the browser sends the query plus locally retrieved chunks; the edge streams a cited response. You only pay for inference when summarization is requested, not for every keystroke.

Why disaggregate?

• No vector DB tax: No Pinecone-style instance running 24/7; the “database” is a static blob on a CDN.
• Local-first privacy: Raw index traffic stays client-side until you explicitly call the edge for AI.
• DX like static hosting: Add a bundler plugin, build, deploy—same mental model as shipping out/ or .next/static.

For package-level APIs and contracts, see Modules. For planned research, see Roadmap.