Zemi

Appearance

Installation ​

Requirements ​

  • Julia 1.12 or later

  • To read papers:

    • Ollama with qwen3-embedding:4b

    • Qdrant on port 6333 (Kaimon can start it)

    • mutool 1.27 or later from MuPDF: apt install mupdf-tools (Ubuntu 26.04, Debian 14 or later), dnf install mupdf (Fedora 43 or later), or brew install mupdf-tools

    • access to a summarizer (see Models)

Without Ollama and Qdrant, the site and keyword search still work; the Status page lists what is missing.

An agent summarizer may need tools of its own. Claude Code reads a PDF too large to attach as page images made by poppler's pdftoppm (apt install poppler-utils or brew install poppler); Zemi itself does not use it. Other agents may prefer other tools: their own documentation covers what they need.

Run ​

sh
git clone https://github.com/kahliburke/Zemi.jl.git
cd Zemi.jl
./zemi

The first run installs dependencies (a few minutes), then serves the site at http://localhost:8000 (ZEMI_PORT to change). Open it from the address zemi prints, which carries a key (?key=…), or with zemi open: the browser keeps the key, and other accounts on the machine, without it, are kept out. The front page leads to Setup: the group and its focus areas, the watched arXiv categories and a contact email (sent to arXiv and OpenAlex), the summarizer, and whether to read alone, join a group's dataset or start one. Setup is later in the top bar's menu.

./zemi help lists the other commands: zemi sync [--unread] brings the knowledge base up to date, zemi bib [FILE] writes every paper, upload and record as BibTeX, and zemi publish, zemi pull and zemi token manage a dataset (see Datasets). pkg> app develop ./app installs zemi on your path.

Add papers ​

From Add documents on the Documents page, an agent's ingest tool, or Julia:

julia
using Zemi
db = kb()
ingest!(db, ["1706.03762", "https://arxiv.org/abs/1412.6980"])
search(db, "how the method's error scales with the problem size")

Ingesting fetches, parses, summarizes and embeds each paper. Finished steps are skipped, so a failed run can be repeated. PDFs not on arXiv and Git repositories (as codebases) are added there too, and the same page finds, reprocesses and drops documents.