Zotero Skimming

Warning

Use the add-on for a fast first pass only. Automated sentence selection can miss important sentences or introduce bias. Do not use the add-on for a detailed study of a paper.

Zotero Skimming selects important sentences in a PDF and saves them as native Zotero highlights. It does not change the PDF.

Skimming is a fast first pass over a paper. It is also called orientation reading.

Highlights stay temporary until you save them. Generate a preview, adjust the threshold, then save the sentences that you want in your library. The reader panel does not open dialogs.

Requirements

• Zotero 9 or 10.
• Optional: Ollama. Ollama runs the local summary and embedding models.
• Optional: an OpenAI-compatible API endpoint and API key. The add-on uses the endpoint for remote summaries and embeddings.

If local summarization fails, the add-on falls back to a built-in TF-IDF (word frequency) ranker. Remote API needs a working endpoint and API key.

Installation

1. Download the latest .xpi file from the GitHub releases.
2. Open Tools → Add-ons in Zotero.
3. Open the gear menu and choose Install Add-on From File….
4. Select the .xpi file. Restart Zotero if Zotero asks.
5. Open Settings → Zotero Skimming and choose a summary source.
6. Optional: install Ollama from https://ollama.com/download.

Ollama runs the models on your computer. If your computer cannot run the models, use a remote API. For Google or Anthropic, use a local LiteLLM gateway. LiteLLM provides an OpenAI-compatible API.

Usage

1. Open the PDF in the Zotero reader.
2. Open the Paper skim section in the right pane.
3. Click Generate highlights. The status line shows extraction, ranking, and model progress.
4. Move the Highlight threshold slider to filter eligible sentences by normalized importance. Zero shows all eligible sentences. 100 hides all eligible sentences. The preview updates without another model run.
5. Click Save as Zotero annotations… to keep the visible highlights in your library. Click Clear to remove the preview.

All other settings live in Settings → Zotero Skimming:

• The summary source. Choose Local Ollama or Remote API.
• The compression ratio controls how many diverse candidates receive priority.
• The highlight threshold. Zero shows all eligible sentences. 100 hides all eligible sentences. The default is 75%.
• Semantic relevance. The add-on then ranks sentences by meaning.
• Classify highlights from labeled summary and the TOML label definitions.
• The endpoint, API key, and model names for remote use.
• The Ollama command, map-reduce, and the number of sentences per chunk.

The Test Ollama button reports its result on the settings page. In remote mode the button checks the endpoint and key fields and does not send a paper.

How it works

Summary

The add-on uses a local, model-free ranker by default. It does not start Ollama or send text to a remote service for this step.

When you enable Semantic relevance or labeled highlights, the add-on uses the selected summary source. The add-on also uses a remote source when you select Remote API.

Map-reduce

Map-reduce (split, summarize, then combine) is optional. The add-on splits long text into chunks and summarizes each chunk. It then summarizes the partial summaries. Local chunks contain N sentences. Remote map-reduce uses a 4096-token context window. The settings pane shows the control in local mode. The setting also applies to remote summaries. Use map-reduce for long papers or small models.

Extraction

The add-on reads the character geometry that the Zotero reader already has. It groups characters into visual lines and separates wide gaps into columns, cells, and labels.

The extractor removes material that is not body prose. It removes printed line numbers, running headers and footers, page numbers, bullet markers, superscripts, footnotes, captions, table cells, and run-in list labels. It also removes titles, authors, headings, abstract text, and reference entries.

Column regions are read in order. Paragraph blocks keep their boundaries, so sentences do not join across columns, captions, or tables. Short sentences remain when they are complete prose. List items remain only when they form complete sentences. Small capitals inside logotypes, such as the LaTeX mark, stay in the text.

The sentence segmenter then divides each block, so each highlight rectangle matches the printed text. The committed corpus in test/fixtures/pdf-extraction measures this behavior.

Ranking and highlights

The ranker scores all eligible sentences from 0 to 1. It gives priority to diverse candidates from its shortlist. The threshold then hides sentences below its value.

Highlights follow the summary sentence order. A position weight keeps the selected sentences spread across the paper. The highlight sequence then explains the same concepts in the same order as the summary. Selected sentences score 1. Other candidates score up to 0.5, so the default threshold shows only the aligned sequence.

Semantic relevance uses embeddings, which are number lists that represent meaning. This option needs a summary and model inference. Labeled highlights also need a summary. With Local Ollama selected and both options off, the add-on selects highlights without models.

The add-on paints the selected sentences as temporary overlays. Nothing is written to the library until you click Save as Zotero annotations…. A second save skips sentences that already have an annotation, so repeated clicks do not create duplicates. Each saved highlight receives the autoskim-key-sentence tag.

With Classify highlights from labeled summary, the add-on compares each selected sentence with the summary labels or the label descriptions in the TOML box. It assigns the closest label and the label color.

scoring-config.json contains the score weights. model-identifiers.json contains the default model names.

Models

Stage	Default model
Summarization	hf.co/unsloth/Qwen3.5-2B-GGUF:Q8_0
Semantic relevance	hf.co/PeterAM4/Qwen3-Embedding-0.6B-GGUF:Q8_0

Test Ollama pulls these models through the Ollama API. Ollama downloads the files from Hugging Face. Enter another model name in Summary model or Embedding model to use a different model.

Similar plugins

SkimRead (repository) provides document-wide skimming for Zotero PDFs. A model selects the sentences. This add-on now follows its settings-pane, temporary-highlight, and explicit-save workflow.

Nodus (repository) is a research workspace for Zotero. It includes search, cited answers, and sentence highlighting.

Build and test

Requirements: Bash, Node.js, Yarn, zip, and git.

yarn install
./scripts/release.sh --only-build
yarn test
node --check bootstrap.js
node --check content/annotator.js
node --check content/extractor.js
node --check content/nlp.js
node --check content/model-manager.js
node --check content/remote-llm.js
git diff --check

./scripts/release.sh --only-build builds zotero-skimming.xpi in the repository root. It reads the version from manifest.json and includes the JSON configuration files, the icon asset, and the locale files.

Without --only-build, the script updates the version, updates updates.json, commits, tags, and can create a GitHub release.

Run the end-to-end smoke test against a real Zotero 10 installation and Ollama:

SMOKE_PDF="/path/to/paper.pdf" \
SMOKE_OLLAMA_COMMAND="mise exec -- ollama" \
yarn smoke

The smoke test runs the working tree in Zotero with a temporary profile. It checks summaries, embeddings, and highlights. It needs Ollama with the two models. See AGENTS.md for the options.

The PDF extraction corpus contains downloaded templates, 16 controlled layout variants, and independent expected text. The test suite tracks extraction quality against a recorded baseline.

For a separate quality report, use Node.js from the repository root.

node scripts/evaluate-extraction.mjs --baseline test/fixtures/pdf-extraction/baseline.json

References

• K. Rayner, E. R. Schotter, M. E. J. Masson, M. C. Potter, and R. Treiman, “So Much to Read, So Little Time,” Psychol Sci Public Interest, vol. 17, no. 1, pp. 4–34, Jan. 2016, doi: 10.1177/1529100615623267.
• R. Fok et al., “Scim: Intelligent Skimming Support for Scientific Papers,” Proceedings of the 28th International Conference on Intelligent User Interfaces. ACM, pp. 476–490, Mar. 27, 2023. doi: 10.1145/3581641.3584034.
• G. B. Duggan and S. J. Payne, “Text skimming: The process and effectiveness of foraging through text under time pressure.,” Journal of Experimental Psychology: Applied, vol. 15, no. 3, pp. 228–242, 2009, doi: 10.1037/a0016995.

Repository

https://github.com/00sapo/zotero-skimming