Zotero Evidence

Zotero Evidence transforms research papers into structured, traceable evidence datasets with AI-assisted screening, extraction, and synthesis.

Define a Codebook, and Zotero Evidence extracts evidence from your literature into a structured Tableau:

Each value remains traceable to its supporting evidence in the original paper.

New: JEV model support — a faster, lighter model pre-screens with confidence scores before you review. See JEV pre-evaluation.

• Import & dedup — import RIS/BibTeX/MEDLINE/PubMed XML; duplicates removed automatically.
• Title/Abstract screening — AI suggests Include/Exclude/Unclear for each paper with reasoning; you confirm.
• JEV pre-evaluation (optional) — a faster model gives a quick, confidence-scored pass before you screen.
• Full-text screening — AI checks each paper against every criterion with highlighted evidence; you confirm.
• Extract coding — AI extracts data into your Codebook, backed by highlighted quotes; you confirm.
• Synthesis — AI groups confirmed evidence into themes with one click.
• Consistency — measure AI-human agreement (Cohen's ϰ), or sample a batch for two reviewers to co-screen.
• Export — PRISMA data, screening log, coding data, synthesis output, and the Codebook itself (CSV).
• Archive & share — export a project as a .zip to back up or share; restore it anywhere.

1. Install the plugin in Zotero 7 or later.
2. File → AI Settings → AI Provider Settings… — set endpoint, model, API key, and concurrency.
3. File → Evidence Project → New Project…, then File → Import Literature → Import to Sources….
4. File → Screening Criteria → define one set of inclusion/exclusion criteria.
5. File → Codebook → define variables (or import from CSV) before coding begins.
6. Work through TA-Screen Queue → FT-Screen Queue → Extract Coding.
7. File → Consistency Calculation → Human-AI Screening Consistency….
8. File → Synthesis Analysis → Theme Mining…, then File → Export Data when ready to write up.
9. File → Evidence Project → Archive Project… to back up or share a project as a .zip.

For a second opinion instead of just AI-vs-human:

1. File → Consistency Calculation → Human-Human Screening Consistency…, pick a sample size, then save the sampled archive.
2. Send that .zip to two reviewers. Each restores it separately and independently screens every sampled item through TA and FT in their own copy, then exports their own screening log.
3. Import both reviewers' CSVs back into the dialog — Cohen's ϰ is computed automatically, and you can start another round any time.
4. Click Apply Agreed Results: items both reviewers agreed on become the project's official result; items they disagreed on stay in TA-Screen Queue, flagged with a colored tag so they're easy to spot in the items list, with both reviewers' calls shown right in the sidebar for a third reviewer to resolve through the normal screening flow.

Repeat steps 1–4 as many times as you like, and every round's own agreement/ϰ stays listed in the dialog.

An optional AI pre-screening pass over TA-Screen Queue, ahead of (or alongside) your own screening — useful for spotting likely-relevant items early and getting their PDFs ready in advance.

1. File → Screening Criteria → JEV Pre-evaluation… — set the JEV endpoint (defaults to aihubmix's /v1/systemone) and API key, and use Test Connection to confirm it's reachable before running a full batch.
2. Click Run JEV Pre-evaluation. Each item gets include/exclude/unclear probabilities and a confidence score; re-running only evaluates items that don't already have a result, unless you check "re-evaluate already-completed items." A running batch can be cancelled at any time, and reopening the dialog mid-run picks its progress back up.
3. Filter by minimum confidence and/or decision to focus on the items that matter most, then fetch full text for them — one at a time, or in bulk for everything currently in view — via Zotero's own full-text retrieval.
4. During Title/Abstract screening, each item's JEV result appears as a clearly labeled reference card. It's informational only: your Include/Exclude/Unclear call is the one that counts.

Built with zotero-plugin-scaffold and zotero-plugin-toolkit.

git clone https://github.com/etShaw-zh/zotero-evidence.git
cd zotero-evidence
cp .env.example .env   # set ZOTERO_PLUGIN_ZOTERO_BIN_PATH and a dev profile
npm install

• npm start — dev server with hot reload
• npm test — run the test suite inside a real Zotero instance
• npm run build — production build (.scaffold/build/)
• npm run lint:check / npm run lint:fix — Prettier + ESLint
• npm run release — bump version and publish

src/
|-- hooks.ts          # lifecycle hooks, File-menu dispatch
|-- modules/
|   |-- project/      # project + Collection structure
|   |-- import/       # Zotero.Translate.Import wrapper
|   |-- dedup/        # DOI-first / title+author+year dedup
|   |-- screening/    # TA judgment, FT per-criterion checklist, criteria, decisions, JEV pre-evaluation
|   |-- consistency/  # Human-AI & human-human screening consistency (Cohen's ϰ)
|   |-- coding/       # Codebook + Extract Coding services
|   |-- synthesis/    # theme mining over confirmed coding evidence
|   |-- pdf/          # text extraction, quote location, highlights
|   |-- ai/           # AI provider config, chat completions, usage tracking, JEV client
|   |-- export/       # PRISMA / screening log / coding export
|   |-- archive/      # project archive export/restore (.zip)
|   |-- db/           # SQLite schema and migrations
|   `-- ui/           # item-pane sections and dialogs
addon/                # manifest, locales, static content
test/                 # Mocha suite, run inside Zotero via `npm test`

Copyright © 2026 Jianjun Xiao. AGPL-3.0-or-later, see LICENSE.