Zotero + local LLM: summarize and query your bibliographie

#Why connect a local AI to Zotero

A researcher quickly accumulates hundreds of papers in Zotero, most of them still unpublished, under embargo, or covered by a confidentiality clause. Pasting the PDF of a manuscript under review into a cloud chatbot means entrusting its contents to a third party—sometimes for training its models. Running an LLM locally solves the problem at its root: the files stay on your disk, inference runs on your machine, and nothing leaves.

The value of using Zotero as a research AI goes beyond privacy. Your library is already structured: collections by project, clean tags and metadata, annotated PDFs. It's a ready-to-use document repository for an LLM. Instead of downloading and reorganizing papers again, you connect the model directly to what Zotero has already organized.

Summarize

Get the objective, method, and results of a twenty-page article in your language in thirty seconds.

Query

Ask a question and let the model search for the answer across an entire collection, not just a single PDF.

Compare

Compare the methods or results from several papers to sketch out the state of the art.

Privacy

Process manuscripts under embargo or collaborator data without a data-leakage clause.

#How Zotero and the LLM communicate

Zotero does not natively “talk” to a language model. The bridge is the files: Zotero stores each PDF on your disk, and that PDF is what you give the LLM to read. Two main approaches coexist, and they are often combined.

On-demand summary

We extract the text from a specific PDF and send it to Ollama with a summarization instruction. Simple, fast, and ideal for one article at a time.

RAG over the library

You index an entire folder of PDFs in a vector database, then query it in natural language. The model reads only the relevant passages, allowing you to cover dozens of papers.

In both cases, the central piece remains Zotero's storage folder, where all the PDFs reside. Zotero 7 also exposes a local API (on port 23119) that lets you list its items programmatically, but to get started, pointing directly to the files is the most robust and transparent approach.

#Prerequisites

Zotero 7

With your PDFs attached to the references (not just the metadata). The built-in PDF reader is not required, but the files must be stored locally.

Ollama

The daemon that serves models, on http://localhost:11434 by default. See the installation guide if you haven't done so yet.

A summarization model

Qwen 3.5 9B (≈6.6 GB of VRAM in Q4_K_M, 256k context, and vision) or Mistral Small 24B for excellent French; a Qwen 3.5 4B (3.4 GB) is sufficient on a small setup.

An embeddings model

For RAG: nomic-embed-text or mxbai-embed-large, retrievable through ollama pull. They are lightweight and run even without a GPU.

An optional RAG component

AnythingLLM or Open WebUI if you want to query your entire bibliography without coding.

# Modèle de synthèse (adaptez à votre VRAM)
ollama pull qwen3.5:9b

# Modèle d'embeddings pour le RAG
ollama pull nomic-embed-text

#Step 1: Find your PDFs in Zotero

Zotero stores each attachment in a subdirectory of the storage directory, named with an eight-character key. You don't need to understand this directory structure in detail: you just need to know where it is so you can point the LLM to it.

Windows

C:\Users\<vous>\Zotero\storage

macOS

~/Zotero/storage

Linux

~/Zotero/storage

If in doubt, the exact path is shown in Zotero under Edit ▸ Settings ▸ Advanced ▸ Files and Folders ▸ “Data Directory.” The storage folder is directly inside it.

To work cleanly on a specific project, it is best to export a collection rather than search through all of storage. Right-click a collection ▸ “Export Collection”: Zotero can copy the PDFs and a bibliography (BibTeX, CSV, JSON) into a dedicated folder, which you can then provide to the LLM.

#Step 2: summarize a scientific article

The most common case: you open a paper in Zotero and want its summary before deciding whether it deserves a full read. We extract the text from the PDF, then send it to Ollama. The pymupdf (fitz) library extracts it cleanly and quickly.

pip install pymupdf requests

import sys, fitz, requests

def extraire_texte(pdf_path):
    doc = fitz.open(pdf_path)
    return "\n".join(page.get_text() for page in doc)

SYSTEM = (
    "Tu es un assistant de recherche. Tu résumes des articles "
    "scientifiques en français, avec rigueur, sans rien inventer. "
    "Tu t'appuies uniquement sur le texte fourni."
)

texte = extraire_texte(sys.argv[1])

resp = requests.post("http://localhost:11434/api/chat", json={
    "model": "qwen3.5:9b",
    "stream": False,
    "options": {"num_ctx": 16384, "temperature": 0.2},
    "messages": [
        {"role": "system", "content": SYSTEM},
        {"role": "user", "content": PROMPT + "\n\n---\n" + texte},
    ],
})

print(resp.json()["message"]["content"])

We call the Ollama REST API on port 11434: it separates the system message (the role) from the content (the article). Two settings matter here — a low temperature (0.2) to limit embellishment, and a sufficiently large num_ctx to ingest an entire paper without truncating it.

#A reliable summarization prompt

Summary quality depends more on the prompt than on the model. A vague prompt produces a generic paragraph; a structured prompt produces a reusable reading brief. The secret is to require a section-based format modeled on the structure of a scientific paper and forbid the model from going beyond the text.

Rédige une fiche de lecture en français de l'article ci-dessous, en respectant EXACTEMENT ce format :

## Question de recherche
Ce que l'article cherche à établir, en une à deux phrases.

## Méthode
Données, protocole, modèles ou outils utilisés.

## Résultats principaux
- Une puce par résultat marquant, chiffré si l'article donne des chiffres.

## Limites annoncées
- Les limites que les auteurs reconnaissent eux-mêmes.

## À retenir pour mon état de l'art
Deux à trois phrases sur l'apport et le positionnement de l'article.

Règles :
- Ne reprends que ce qui est réellement écrit dans l'article.
- Si une section n'est pas renseignée dans le texte, écris « non précisé ».
- N'invente aucun chiffre, aucune référence, aucun nom d'auteur.

The imposed format

Section headings force the model to sort instead of rambling. You get the same framework from one article to the next, comparable at a glance.

The anti-hallucination rule

“Only repeat what is written” and “not specified” reduce the risk that the LLM will invent a result or reference—the main danger in scientific contexts.

The state-of-the-art section

By explicitly requesting the positioning, you turn the summary into raw material that can be directly reused in a literature review.

#Step 3: query your entire bibliography with RAG

Summarizing a PDF is useful; querying fifty papers at once is what changes how you prepare a literature review. That is where RAG (Retrieval-Augmented Generation) comes in: you split PDFs into chunks, convert them into vectors with the embedding model, and the LLM reads only the passages relevant to each question.

The simplest approach, without writing code, is to point AnythingLLM or Open WebUI to the folder exported from Zotero (or directly to storage). These tools handle indexing and the vector database for you; you only need to choose Ollama as the provider and nomic-embed-text as the embeddings model.

1. 01
   Export the collection

   From Zotero, export the relevant collection with its PDFs to a dedicated folder, such as ~/recherche/etat-de-l'art.
2. 02
   Create a workspace

   In AnythingLLM, create a workspace and set the LLM provider to Ollama (http://localhost:11434) and the embedding model to nomic-embed-text.
3. 03
   Import documents

   Drag the PDF folder into the workspace. The tool automatically extracts, chunks, and indexes the text.
4. 04
   Query in natural language

   Ask cross-cutting questions: “Which evaluation methods come up most often?”, “Which papers contradict hypothesis X?”

RAG’s main advantage over simple summarization: the model cites the passages it used. You can trace them back to the source PDF to verify them, which is essential in research. Always ask the model to indicate which document each claim comes from.

#The limitations on highly technical papers

This is where you need to be honest: in a math, theoretical physics, or highly formal ML article, a local LLM quickly shows its limitations. Understanding these blind spots helps you avoid trusting it in the wrong place.

The formulas

Text extraction flattens equations: subscripts, exponents, and Greek symbols get mixed up or disappear. The model then reasons over corrupted formulas and may draw false conclusions.

The tables

A table turns into a jumble of numbers when extracted as plain text. Row/column relationships are lost, so figures cited from a table are unreliable.

The figures

A text model cannot see charts. Anything that exists only in a figure is completely lost to it—it won't mention it, or it will make it up from the caption.

Mathematical reasoning

Following a demonstration or checking a derivation exceeds what a local 14B model can do reliably. It paraphrases the proof without validating it.

In practice: use the LLM for the “narrative” layer of a paper — research question, motivation, stated contributions, limitations, and positioning. Keep human oversight for every formula, numerical table, and figure. For papers rich in equations, a multimodal model that can read pages as images rather than extracted text delivers better results, but it remains a supplement, not a substitute for your own reading.

#Troubleshooting

The PDF comes out blank

This is a scan without a text layer. Run it through OCR (ocrmypdf entree.pdf sortie.pdf) before extraction.

The summary skips the end of the article

num_ctx is too small: the text was truncated. Increase it if VRAM allows; otherwise, summarize by section and then summarize the summaries.

RAG gives irrelevant answers

The wrong passages were retrieved. Rephrase the question using the domain's exact terms, or reduce chunk size during indexing.

Slow responses

A 9B model crawls on the CPU. Make sure Ollama is actually using the GPU, or switch to a Qwen 3.5 4B (3.4 GB) for routine summarization.

Numbers that don't match the PDF

A classic symptom of a poorly extracted table or a hallucination. Lower the temperature and, above all, cross-check against the source.

#Go further

This guide builds on components already covered in detail on the site. To explore installation or RAG in particular:

Install Ollama: Windows, macOS, and Linux

The starting point for serving your models locally on port 11434, if you haven't set it up yet.

Local RAG with Ollama without coding (Open WebUI, AnythingLLM)

The definitive guide to building the RAG component that queries your entire bibliography without writing a line of code.

Local RAG with LM Studio: chat with your documents

An all-in-one alternative if you prefer LM Studio to the Ollama + separate interface pair.