EvilFlowersCatalog 0.15.0
EvilFlowersCatalog 0.15.0 was released on 12 November 2026. Search now runs inside the catalog's own PostgreSQL, and the catalog can search inside its publications and answer questions about them with page citations. After a round of load testing, the list endpoints also run far fewer database queries.
This post summarises the release for the people who use and run the catalog. The full release notes list every change with its issue.
For readers
- Better search. Full-text search with Slovak and Czech stemming, so a
query also finds other forms of the same word. Hits are highlighted, a
misspelled query gets a "did you mean", and a new endpoint suggests titles as
you type (
GET /api/v1/entries/suggest). - Facets with exact counts for filtering results, in the API and in OPDS 1.2 and 2.0 search feeds, so e-reader apps get them too.
- Ratings and reviews on publications, with moderation.
- Reading history of what you opened.
- Category and feed titles in English and Slovak.
For librarians
- Lending statistics for the admin dashboard.
- Branded notification e-mails, and e-mail addresses imported from LDAP.
- Publishing from AI agents. The MCP server gained tools to create publications and upload their files, and to manage users and API keys.
- Readers can change their own password.
- Fixes: one-week loans can be renewed, returning a loan that was never opened sends its e-mail, and empty uploads are refused.
Search by meaning, extraction and the assistant
These parts are in the release. Each works only once the service behind it is configured.
- Ranking by meaning. With an embedding service configured, search combines word ranking (BM25) with semantic ranking. Without one, it ranks on words only.
- Extraction is off by default (
EVILFLOWERS_EXTRACTION_ENABLED). The catalog stores paragraphs, figures and their embeddings produced by the extraction workers. That enables search inside publications (GET /api/v1/entries/passages), a publication's figures, and similar publications. - The assistant answers questions from the library's passages, labels how strongly each passage supports the answer, and links to the pages it used. In the PDF reader it explains a selected passage or figure. It talks to a model through the Ollama API, so you can host the model yourself. Without a model endpoint configured, it cannot answer.
- Titles lent with Readium LCP are never read by AI.
For people who run it
- List endpoints cost the same number of queries for any page size.
GET /api/v1/entries?limit=30used to run 226 SELECTs. - The signing key is loaded once per process instead of on every request, which had cost about a third of the CPU.
- Covers and thumbnails have versioned URLs, so browsers and proxies cache them and a repeat request answers 304 without touching storage.
- Downloads support HTTP range requests, so large files stream.
- OpenTelemetry traces, metrics and JSON logs, and optional profiling.
- Optional connection pooling, PgBouncer transaction pooling, and Celery settings for autoscaled workers.
manage.py backupworks with PostgreSQL 18 and fails loudly when the dump fails.
Before you upgrade
Read the upgrade notes at the end of the release. In short:
- PostgreSQL needs the
vectorandpg_trgmextensions.pg_searchis optional and enables BM25 ranking. Our PostgreSQL image has all three. - CORS is no longer open to every origin. Set
CORS_ALLOWED_ORIGINS, or rely on the default, which allows the origin ofEVILFLOWERS_PORTAL_URL. - The MCP server accepts only Bearer API keys by default. Password clients need
EVILFLOWERS_MCP_AUTHENTICATION_SCHEMAS=Bearer,Basic. - The integration with the separate search service is removed. Search runs in
the catalog's database, and
evilflowers-search-serviceis archived. - After migrating, run
manage.py embed_entries --missingif you configured an embedding service.