An RDF graph in one file. Put it on any static host and query it with SPARQL, without a database server.
Try it in your browser · Documentation · Changelog
rete (Latin for "net", pronounced RAY-teh) stores a whole RDF graph in one
immutable .rete file: the term dictionary, the sorted triple indexes, and
optional extras such as a Dataset Card or a full-text index. Put the file on
any host that serves HTTP range requests (S3, Cloudflare R2, GitHub Pages,
Zenodo) and clients run SPARQL against it where it sits. A query fetches only
the parts of the file it needs, so there is no server to run and no import
step. The same engine runs in the CLI, in Python and JavaScript, and in the
browser through WebAssembly.
If you know Parquet or PMTiles, it is the same idea applied to RDF graphs.
Each file can also describe itself. Its Dataset Card (title, license, counts, vocabularies, starter queries) and its schema summary are read in a couple of small range requests, without touching the indexes.
rete is built for graphs that you publish and other people read: a dataset release, a nightly dump, a knowledge-base export, one file per tenant or per year. You host a file instead of running a triplestore, and people query it from a browser tab, a notebook or a shell.
It is not a database. Files are immutable, so a change means a rebuild. If you
need writes, transactions or an always-on endpoint under load, use a
triplestore such as Oxigraph, Jena or
GraphDB. There is no full OWL/RDFS entailment at query time either: rete
rewrites queries for OWL 2 QL (--entail) and can bake RDFS/OWL RL inferences
into the file at build time (--materialize); see
reasoning.
cargo install rete-cli --locked # the rete command-line tool
pip install rete-graph # Python 3.9+, also in Pyodide
npm install rete-graph # JavaScript: Node 18+ and browsersNo Rust toolchain? The CLI is also published as a small container image,
ghcr.io/caviri/rete-cli (amd64 and arm64). With this alias the commands
below work as written on the files in your current directory:
alias rete='docker run --rm -i -v "$PWD:/data" -w /data ghcr.io/caviri/rete-cli:latest'Read the three container pitfalls in
Getting started
first: piping needs -i, the image runs as root on Linux, and Git Bash on
Windows rewrites paths.
Build a file from an RDF dump. N-Triples, N-Quads, Turtle, TriG, RDF/XML and
Jelly are detected by extension, and several inputs merge into one file.
--card embeds a Dataset Card.
rete build dump.nt -o graph.rete --card --title "My graph"Inspect and query it locally:
rete card graph.rete
rete sparql graph.rete 'SELECT ?s ?name WHERE { ?s <http://xmlns.com/foaf/0.1/name> ?name } LIMIT 10'A remote file works the same way, with nothing to download first. OpenCitations Meta is a single 36 GB file with 5.2 billion triples; reading its card takes two small requests:
rete card-url https://data.graphplaza.com/opencitations/opencitations.rete
# fetched 1740 of 35852509508 bytes in 2 range request(s) — index NOT fetched
rete sparql-url https://data.graphplaza.com/opencitations/opencitations.rete \
'SELECT * WHERE { ?s ?p ?o } LIMIT 3'The CLI reference covers every command, and hosting explains how to put your own file online.
To try it without installing anything, open the playground. It runs the same engine in your browser, on example datasets that range from small graphs embedded in the page to remote files with billions of triples.
All of these run the same engine.
- CLI and Rust: the
rete-clibinary, or therete-corecrate as a library (CLI, Rust API). - Python:
rete-graphon PyPI, for CPython and Pyodide (JupyterLite, marimo) (docs). - JavaScript:
rete-graphon npm, for Node and browsers, with TypeScript types (docs). - R: installed from GitHub with
remotes(needs Rust); not on CRAN (docs). - Java: pure JVM with an RDF4J binding, built from
clients/java; not on Maven Central yet. - Browser, no install: the playground and a SPARQL IDE query any
.reteURL. - Claude Desktop: download rete.mcpb and double-click it; the engine runs on your machine (docs).
- Blender: an add-on from the releases page that turns query results into 3D scenes (docs).
- Hosted gateway: katospiegel-rete.hf.space serves MCP, a REST API and a SPARQL endpoint for each published dataset.
This repository is also a Claude Code plugin. It adds skills for building and
publishing .rete files, and the rete MCP server for querying the public
datasets:
/plugin marketplace add caviri/rete
/plugin install rete-graph@rete
The same MCP server works in other MCP clients, ChatGPT included; see agentic interfaces.
A 1 KiB header points to every section, so a reader fetches the header and then only the sections a query needs. Optional sections a file does not carry are not listed. The format spec and architecture notes have the details.
The indexes are built into the file, so opening it needs no parsing or indexing, and aggregates stream: counting the 779,399 software records in the 9.83-billion-triple DataCite file takes 4 s inside a 2 GiB container (benchmark).
- Quickstart: build a small graph and understand its first query.
- Build and publish or query and analyze: choose a task and follow its guide.
- Technical reference: exact commands, APIs, format details and compatibility limits.
- Worked examples and apps: complete workflows, notebooks and visual demonstrations.
rete is pre-1.0, and both the API and the file format can still change. The
format (byte 0x05, stable generation 1) has not changed since it was frozen
in July 2026, but that is a track record, not a promise: until 1.0.0 a release
may require you to rebuild files you have published, so keep the RDF you built
them from. See compatibility.
New in 0.4.0 for people who build files: a label index (--label-index) and a
schema section that no longer depend on the pyramid, --text-index in
memory-bounded builds, an opt-in quad index for named graphs
(--named-graphs quad, which older readers refuse), and rete recard, which
replaces a file's Dataset Card without a rebuild. The
CHANGELOG has the rest.
Known limits:
- Reads are lazy, but evaluation is not fully streaming. A query fetches only
the tiles and dictionary chunks it touches, while hash-join build sides,
GROUP BYgroups,DISTINCTsets and anORDER BYwithoutLIMITare held in memory. A small local file is read whole. Details. rete federateunions the results of several files; it does not join across them. The playground can split a basic graph pattern across files, with limits. Details.- The browser engine is single-threaded by default; multi-threading is experimental. Details.
Everything builds in Docker, nothing on the host:
docker compose build dev
docker compose run --rm dev cargo build --release -p rete-cli
docker compose run --rm dev cargo test --workspace --exclude rete-bench
docker compose run --rm dev bash scripts/smoke.sh
bash tests/gate/gate.sh # browser regression gateYou can also open the folder in its devcontainer. CONTRIBUTING.md and AGENTS.md list the full set of checks and the rules for generated files.
rete is free and open source. If it is useful to you, you can support its development on Ko-fi:
