Observed arrival · 2026-11-13
Atrium keeps the page attached to the chunk
A document-ingestion library that turns PDFs, Office files, HTML, EPUB, and scans into structure-aware chunks with provenance, then serves retrieval over HTTP.
- For
- Developers building document search and retrieval systems
- Worth noticing
- The reported 88 MB serving footprint excludes the separate Ollama embedder, measured at about 670 MB.
Field notes
The documented flow posts a file to a local /ingest endpoint, parses it on the CPU, then hands content to a separately running embedder; the page names Ollama and bge-large. Chunks retain page, heading path, element type, and the strategy used, while the demo compares five strategies and notes that “auto” currently resolves to structure_aware. On an Apple M4, the project reports 27/36 ms p50/p95 search over 1,451 chunks, excluding embedder memory.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue