Chat with any PDF open in Zotero's reader, grounded in the paper's own tables, figures, equations, notes, and bibliography.

---

LLMz (pronounced "el-el-em-zee", /ˌɛlˌɛlˌɛmˈziː/) is an experimental Zotero 9 item-pane plugin that turns your Zotero into a research assistant by adding an LLM chat pane next to the reader. Ask a question about the paper you're reading and it automatically pulls in whichever paragraphs, tables, figures, equations, your own highlights/notes, and bibliography entries are actually relevant, then answers with clickable citations that jump straight to the right spot in the PDF.

The goal of this personal project was to quickly develop a tool that could mimic some of the functions of NotebookLM, while integrating with Zotero and allowing for the possibility to be run fully locally.

Disclaimer: LLMz is an independent, community-built project. It is not affiliated with, endorsed by, or sponsored by Zotero or the Corporation for Digital Scholarship.

Trailer

LLMz.mp4

Features

• 🤝 Multi-provider — Ollama, LM Studio, and LiteLLM for local/self-hosted models, plus OpenAI and Anthropic directly. Switch providers and models at any time; vision support is detected per model. Additional model providers can be provided through a LiteLLM proxy.

  

• 📄 PDF-grounded chat + clickable cross-library citations — Text, tables, figures, equations, your own annotations, and the paper's bibliography are extracted automatically and offered to the model as context, with a single tool-calling round-trip deciding what's actually relevant to your question. The model's answer links back to text excerpts, tables, equations, notes, and page numbers; clicking one jumps the reader to that exact spot. When a question cannot be answered from the active PDF, LLMz pulls knowledge from other documents in your library and references those through clickable links as well.

  

• ⏱️ Conversation history — every chat can be saved as markdown per PDF, browsable, exportable, and importable, allowing you to export conversations for sharing with colleagus, Obsidian, or for resuming conversations later.

  

• 🏞️ Image paste — attach a screenshot or clipping directly into a chat turn.

  

• 🎹 Keyboard-driven — shortcuts for submit, stop, and history navigation (see the in-pane "Keyboard Shortcuts" panel).

• 📚 Reference tools — ask in plain language to download a bibliography entry into your library (with a PDF attached when one can be found), or link it to an item you already have.

• ⬇️ Table export — pull one, several, or all of a paper's tables out as CSV files bundled into a zip, image-grounded when the model supports vision.

Usage

Just type a question about the paper in the Prompt chat box. Depending on what you ask, LLMz will:

• Pull in matching paragraphs, tables, figures, equations, notes, and/or the bibliography as context before answering.
• Recognize a request to download or link a specific reference ("download reference 12 into my library", "link everything by Smith to my library").
• Recognize a request to export tables ("export table 3 as CSV", "export all tables").

Click any text / table / figure / equation / reference / page-number link in a response to jump to that spot in the reader. Use the Export button to save the active conversation for later or to share it with colleagues, to store it in your Obsidian vault or to store it for future LLM ingestion. Click Load to resume a past conversation from this paper's conversation history, or Import to load any conversation shared as .md file.

Requirements

• Zotero 9.
• A running LLM backend Ollama or LM Studio locally, a LiteLLM proxy, or an OpenAI/Anthropic API key.
• Node.js, for the document-structure pipeline (sdt/document-worker/) used to extract equations, references, and notes. npm install both installs its own runtime dependencies and fetches Zotero's document-worker, pdf.js, and structured-document-text sources at pinned commits directly from their own repos (not bundled in this repo -- see Licensing below and PINNED-SOURCES.md):

  cd sdt/document-worker
  npm install

  npm install resolves the correct native binary for your platform automatically (macOS/Windows/Linux, x64/arm64) -- no manual steps needed beyond running it and having git and network access available.
• A Python 3 virtual environment at $ZOTERO_HOME/LLMz/venv, used to extract and crop figures using pymupdf, and to communiate with the sqlite-vec embeddings database, with dependencies specified in requirements.txt:

  python3 -m venv ~/Zotero/LLMz/venv
  ~/Zotero/LLMz/venv/bin/pip install -r requirements.txt

Installation

1. Download the latest llmz.xpi, or build one yourself from source:

   mkdir build
   zip -r build/llmz.xpi . -x ".*" -x "*.xpi" -x "scripts/__pycache__/*"

2. In Zotero, go to Tools → Add-ons, click the gear icon, choose Install Add-on From File..., and select the .xpi.
3. Restart Zotero.

Setup

1. Open Preferences → LLMz (or the pane's own Providers panel) and pick a provider.

   • For a local server (Ollama/LM Studio/LiteLLM), set its host/port if it isn't running on the default.
   • For OpenAI/Anthropic, enter an API key.

2. Open Embeddings and select the provider and embedding model that you want to use. ⚠️ Ideally, this will only be done ONCE. Every time you switch the embedding model, you will still be able to answer questions about the active PDF, but you will have to re-index your library to support cross-library references.

   

3. Open a PDF in the reader — the LLMz pane appears in the item pane alongside it.

4. Pick a model from the dropdown and start asking questions.

5. To support cross-library citations, open up the Library Index panel and click on the Index All button to mine and index all PDF documents in your library. Depending on the embeddings provider and the number of papers in your library, this could take a while.

   

Recommended settings: The plugin has been mostly tested with OpenAI GPT 5.4 as chat model, and using OpenAI text-embedding-3-small as embeddings model, ensuring high response and retrieval quality at an acceptable cost.

How it works

When you open a PDF, LLMz runs it through structured-document-text, which classifies the page layout and returns the document's structure -- paragraphs, tables, figures, equations, and bibliography entries, each anchored to a location in the PDF. pymupdf is then used to render pages and crop out the figures and tables identified by that structure. Together these give the model a set of addressable, citable chunks instead of a flat text dump. When you ask a question, a single tool-calling round-trip lets the model pull in whichever chunks are actually relevant, and its answer links back to them so you can jump straight to that spot in the reader.

Alongside this structural extraction, LLMz embeds every paragraph, table, figure, and equation into a sqlite-vec vector database. These embeddings serve two purposes: they're used to re-ground a citation when the chat model's own reference to a location can't be resolved directly, and, since embeddings are built up across the whole library rather than just the active PDF, they let LLMz pull in relevant paragraphs, equations, figures, and tables from other papers whenever the active document doesn't contain enough context to answer a question on its own.

Navigating to a citation, and gathering context about what you're currently looking at (the active page, a text selection, an annotation), both go through Zotero's own reader API rather than a separate PDF viewer layer. That's what lets a clicked citation jump the actual reader to the right spot, and lets a question implicitly refer to "this page" or "what I've selected" and have LLMz resolve it correctly.

Project structure

llmz/
├── core/                  Plugin logic
│   ├── document/          PDF content extraction: tables, figures, equations, references, notes
│   ├── ui/                Chat pane UI components (chat log, provider settings, history, etc.)
│   ├── llm/               Logic for prompt composition, embedding generation, and for talking to
│   │                      the model endpoints.
│   ├── chat-pane.js       Main plugin entry point (item pane registration, onRender)
│   ├── citation.js        Citation-index building and grounding logic
│   ├── conversation-history.js, semantic-history.js
│   │                      Per-PDF chat history, saved as markdown and semantically searchable
│   ├── import.js, export.js
│   │                      Reference download/link and table export orchestration
│   └── python-setup.js    Bootstraps the venv used for figure extraction
├── tools/                 Native tool-calling features: reference download/link, table export
├── res/
│   ├── icons/             SVG icons
│   └── img/               LLMz logo and README screenshots
├── styles/                Pane stylesheet
├── sdt/document-worker/   The ML-based PDF layout classification pipeline
└── scripts/               Python/Node extraction scripts used to mine the PDF document

Contributing

Feel free to get in touch if you want to contribute to this project, or to open a PR if there is something you would like to see included.

• Email: luismherrmann@gmail.com
• LinkedIn: https://www.linkedin.com/in/luis-herrmann-32468814a/

License

MIT © 2026 Luis Herrmann

LLMz depends on a few pieces of code it does not own the license to, none of which are bundled in this repo -- each is installed or fetched separately at install time, directly from its own source, so it doesn't affect LLMz's own MIT license:

• pymupdf, used for rendering and figure extraction (see Requirements), is dual-licensed AGPL-3.0 / commercial by Artifex. It's installed into your own Python virtual environment, not bundled with LLMz.
• document-worker and pdf.js (Zotero's fork), used by the document-structure pipeline, are licensed AGPL-3.0 and Apache-2.0 respectively. npm install fetches them directly from zotero/document-worker and zotero/pdf.js at pinned commits rather than committing their source here -- see PINNED-SOURCES.md.
• structured-document-text, also fetched by npm install from zotero/structured-document-text, publishes no LICENSE file upstream as of the pinned commit. It's fetched on the assumption that it's licensed the same as its sibling Zotero repositories (AGPL-3.0), pending confirmation -- see structured-document-text/NOTICE.md, written into sdt/document-worker/ by fetch-sources.js once fetched (not committed here, same as the fetched source itself).
• marked, highlight.js, and KaTeX, bundled directly as minified builds in vendor/ -- all permissively licensed (MIT / BSD-3-Clause), compatible with LLMz's own MIT license. See vendor/README.md for versions and per-file details.

None of this is legal advice -- if you're redistributing a packaged install or using LLMz commercially, check the terms of these dependencies yourself.