TinySearch

TinySearch 0.4.0 makes local-first, token-efficient web research easier to
install, run, and integrate. Search, crawling, embeddings, and hybrid reranking
turn webpages into compact, source-grounded evidence without requiring a paid
search API.

What changed

• Installable package: TinySearch is now published as tinysuite-search
  with a proper Python API and optional [server] dependencies.
• One-command MCP: launch over stdio with uvx, or use
  tinysearch serve for Streamable HTTP.
• Setup tooling: tinysearch setup prepares Chromium and the local ONNX
  model; tinysearch doctor checks runtime readiness.
• Unified configuration: one validated tinysearch_config.json controls
  search backends, embeddings, crawling, concurrency, ranking, and token
  budgets.
• Structured evidence: Python, MCP, and FastAPI now share the same research
  and scraping pipelines, source attribution, and stable result schema.
• Better retrieval: search-result ranking, cross-source chunk ranking,
  deduplication, and source quotas keep evidence focused and reduce context.
• Flexible search: use DDGS by default, self-hosted SearXNG, or an optional
  Brave fallback.
• Stronger releases: wheels and Docker images are smoke-tested across
  Python 3.12–3.14, with patched pypdf, MCP, and Crawl4AI dependencies.

Upgrade notes

The package is now named tinysuite-search, server dependencies use the
[server] extra, and the primary configuration file is
tinysearch_config.json. Refresh cached MCP tool schemas after upgrading.

Full changelog: https://github.com/MarcellM01/TinySearch/compare/v0.3.2...v0.4.0