Skip to content

zsum mode

zsum mode runs the zotero-summarizer pipeline: it reads your Better BibTeX export, extracts text from your PDFs, writes a structured card per paper and a digest per collection, and publishes them as an interlinked Obsidian vault.

It needs only the core install and a vault_path.

Tabs

Tab What it does
Overview Pipeline status at a glance
Pipeline Run stages (extract, skeleton, cards, digests, publish, all, stats, clear-cache) with live log streaming
Vault Read and search the published vault, with wikilink resolution and LaTeX math; dot-folders such as .github are hidden
Collections Browse the collection hierarchy with digest state and per-paper card state
Prompts Edit the prompts the pipeline uses; your edits are kept separately from the bundled defaults
Runs Inspect and rerun past pipeline runs
Settings zsum options and the zsum LLM provider override. If the zsum config file doesn't exist yet, Settings offers to create a starter one

Extraction modes

Set extraction.mode in your zsum config:

Mode Behaviour
pages (default) First N and last M pages of each PDF
sections Table-of-contents and heading-detected sections; falls back to page ranges when too few are found
notebooklm Hands card and digest generation to NotebookLM; papers it fails on fall back to page ranges

Commands outside the UI

The bundled zotero-summarizer also has commands that are not buttons yet. They augment an existing vault without a full rerun:

Command What it does
zsum citations Citation counts from OpenAlex, with Semantic Scholar as fallback, cached for 30 days by default
zsum reviews Marks review, survey, meta-analysis and tutorial papers
zsum restructure Suggests renames, splits and merges for your collection hierarchy; never moves papers
zsum nlm upload\|cards\|digests\|all\|status\|cleanup The NotebookLM pipeline, step by step
zsum setup-model --size ... Configures a local Ollama model

What you get

The published vault has a library index, a note per collection with its digest, a note per paper with its structured summary, and wikilinks between them.

Search backends

Vault search uses ripgrep by default. Set search_backend = "qmd" and vault_qmd_collection to use a QMD collection instead.