ISWC 2026 · Posters & Demos · Bari, Italy
Publishing an RDF graph for interactive use normally means running a SPARQL endpoint or shipping the whole dataset to whoever asks. This demonstration takes a third route: the graph, its dictionary, its six permutation indexes and its own description are packed into one immutable file whose sections are addressed by byte range. Put that file on ordinary static hosting and a browser can query it in place — reading only the bytes the query touches.
Three graphs are published here, up to 60 GB and 3.8 billion triples. Nothing behind them is a server. Every station below shows you the answer and what the answer physically cost.
A .rete file opens with a 1 KB header holding a typed section
directory — the offset and length of everything else. Nothing later in the
file is ever found by scanning, so every read is a bounded HTTP
Range request.
The dataset card and the schema pyramid sit before the dictionary, so a visitor can learn what a graph contains — title, licence, provenance, exact per-predicate and per-class histograms, runnable starter queries — without opening a triple index at all. Because all six permutations exist, a triple pattern with bound terms resolves to a contiguous range with its bound components leading, which is what keeps remote routing bounded.
All three are plain objects on Cloudflare R2 behind a CDN — 206 Partial Content, CORS, no token, no redirect. The same URLs feed all three stations.
| Graph | Triples | File | Licence | Source |
|---|---|---|---|---|
crossref
The complete Crossref March 2026 citation graph: 179.5M registered works and 2.0B cites edges — the largest single .rete published. |
3,777,727,303 | 60.2 GB | CC BY 4.0 | Public Data File |
| zenodo-records Every published Zenodo record (7.76M) as a DataCite scholarly graph, with ORCID creators and version chains. | 215,396,999 | 2.09 GB | CC-BY-4.0 | Zenodo exporter |
| open-pulse The EPFL/SDSC research-software graph — repositories, people, labs and their ROR institutions. A third-party graph, not ours. | 3,697,829 | 49 MB | GitHub metadata · ontology © SDSC-ORDES | Open Pulse |
They share canonical IRIs — a work is https://doi.org/…, a
person is https://orcid.org/… — so Crossref and Zenodo join on
DOIs and ORCIDs without any owl:sameAs reconciliation.
Because the file is the interface, anything that can issue an HTTP range request can query it. None of these three has a server behind it.
A static page with the engine compiled to WebAssembly. Open a graph you did not build, read its card, run a summary-safe query that never touches the index, then a selective one — and watch the run strip report the requests issued, the bytes moved, the permutation chosen and the exact byte ranges read.
The same URLs, from Python — no kernel, no server, no local install.
Pyodide runs in the page and rete-graph installs from
PyPI; results land in pandas, and stats() keeps counting
the bytes and range requests each step actually spent.
A packaged MCP extension puts the same engine inside an MCP-capable assistant: nine tools over these files, among them the card, the schema and SPARQL. It reads the card to learn the vocabulary before querying — which is what stops a model inventing predicate names.
Measured on 2026-07-24 against the public files, over a residential link,
end to end — process start and TLS included. Reproducible with the shipped
rete card-url, sparql-url, why-url
and cost commands.
Interaction on zenodo-records.rete |
Requests | Bytes read | % of file | Index read | Time |
|---|---|---|---|---|---|
| Dataset card | 3 | 49,452 | 0.002% | no | 0.7 s |
| Predicate histogram, exact | 3 | 49,452 | 0.002% | no | 0.7 s |
| Selective pattern, 19 triples | 41 | 7,602,176 | 0.36% | partial | 5.7 s |
| Join: version chain → title | 30 | 17,825,792 | 0.85% | partial | 4.3 s |
| Join: ORCID author → works | 50 | 18,743,296 | 0.89% | partial | 7.3 s |
| Join: cited-by, empty result | 24 | 4,980,736 | 0.24% | partial | 3.4 s |
The card's 36 per-predicate counts sum to exactly 215,396,999 — the histogram is exact, not sampled. And the entry cost barely grows with the graph: the same three requests return 2.6–2.8 KB on the 17.5 GB ORCID and 35.9 GB OpenCitations files.
Where the model stops paying off. Bounded, kilobyte-scale access is guaranteed for summary-safe questions and typical for selectively routed ones. An unselective aggregate over a whole graph must read the index and will move a large fraction of the file — the demonstration lets you trigger that too, and watch it happen.
Snapshots are also immutable: updates mean a rebuild. This is a publication format, not a transactional one.
.rete: Serverless SPARQL over Range-Addressable RDF Snapshots. ISWC 2026 Posters & Demonstrations Track, Bari, Italy, 25–29 October 2026.
Read the paper (PDF) Source code (Apache-2.0) Full project site
This site is a frozen snapshot built for the demonstration: three graphs, one console, one notebook, one extension. The live project — 80+ published graphs, the CLI, and clients for JavaScript, Python, R, Java and Blender — continues at caviri.github.io/rete.