Skip to the text
Scholaris

Changelog

What has changed in Scholaris, from the first version to the second, told by milestones and in the order they were made.

Reviewed on This page as Markdown

October 2026

  • 6 October: scholaris.joseluissaorin.com starts serving the second version of Scholaris, on Cloudflare. The first version is kept in reserve.
  • This knowledge base, with a Markdown twin of every page, /llms.txt, /llms-full.txt and the page for agents.
  • The Python SDK, ready to be published on PyPI as scholaris-sdk, under the EUPL-1.2.
  • Exporting an .spdf from the server, with a 24 MB cap on embedded files and a notice of what is left out.
  • Redoing a document's record when the evidence of its identity changes (another episode, another guest), with a dry-run mode.
  • Redoing only a document's figures, with the estimated cost shown first.

The second version, in the order it was made

The second version of Scholaris was written anew. Its milestones, first to last:

  1. Foundations: the SPDF 4.0 format and the imprenta, the converter that reads any file in the browser.
  2. Search by three fused routes (lexical, semantic and visual) and answers with validated citations; the official CSL styles.
  3. Reading pages in batches, printed folios, records, fragments, vectors and figures; the concept map.
  4. Citations: temporal logic, autocite with a judge, .docx insertion and BibTeX.
  5. The API on Cloudflare Workers, with Clerk sign-in and encrypted own keys; SPDF on Workers and migration of old SPDFs.
  6. The web app (library, reader, search, writing, exploring), the home version on Node and SQLite, the desktop executable and the Docker image.
  7. A new reader, Gemini 3.8 Flash, chosen with the benchmark.
  8. Media: YouTube by address, podcasts, Vimeo and direct links.
  9. The MCP server with OAuth 2.1 and migration of first-version accounts.
  10. Reading in three stages (readable, searchable, vectorised), economy mode and the entity graph linked to Wikidata.
  11. The quality benchmark: nDCG@10 of 0.889 and no invented citations.
  12. Two-stage search and the new player, highlighting the word being spoken.
  13. API v1 with its guide at /en/api, the multi-user home version and shared libraries.
  14. Drawings and motion in the app; records for radio and TV programmes; OpenAlex with a cache.
  15. Getting ready for production.

The first version

The first Scholaris ran on a home server and stored documents as SPDF 1 to 3. Its SPDFs still open: they are migrated to format 4 on the fly. Compared with it, the second reads a 54-minute interview in 45 seconds instead of three hours and 39 minutes, and finds better (nDCG@10 from 0.695 to 0.889). See Performance.