Jew.tech → Torah Graph

The Torah knowledge graph

A knowledge graph of Jewish text and tradition, built from CC-BY Sefaria sources: 157,143 nodes and 215,426 relationships in which rabbinic teachings carry their Hebrew, their citable reference and their vector embedding, verses are first-class objects, and the chain of who-taught-whom is machine-readable. Every number on this page was re-verified against the live database on 2 August 2026, and every caveat that verification turned up is published beside it.

The graph runs on Neo4j 5.26 in a container on a single machine. There is no cloud service behind it and no vendor to take on faith: the counts above come from MATCH (n) RETURN count(n)-class queries whose exact text is kept in a claims ledger alongside the result and the date, and they are re-run before anything is published. What follows is the graph's actual shape — including the parts that are messy.

What a cited teaching looks like

The flagship object in the graph is the teaching. There are 17,147 of them, and the coverage is total rather than typical: every single one carries a citable Sefaria reference, the Hebrew source text, the English, extracted themes, a license, and a 768-dimensional embedding for semantic retrieval. Nothing in this graph can be quoted without its source travelling with it. Here is one node, printed in full except for the embedding:

ref

Pirkei Avot 1:1

work

Mishnah Avot (Sefaria)

source

Mishnah, Avot

url

https://www.sefaria.org/Pirkei_Avot_1.1

license

CC-BY

categories

Mishnah

themes

learning, humility, deeds, character, love, justice

embedding

768 floats (omitted here)

he — Hebrew source text

משֶׁה קִבֵּל תּוֹרָה מִסִּינַי, וּמְסָרָהּ לִיהוֹשֻׁעַ…

text — English

Moses received the Torah at Sinai and transmitted it to Joshua, Joshua to the elders, and the elders to the prophets, and the prophets to the Men of the Great Assembly. They said three things: Be patient in [the administration of] justice, raise many disciples and make a fence round the Torah.

That first sentence is not decoration. Moses received the Torah at Sinai and transmitted it to Joshua is a claim about a graph — a directed edge, asserted at the head of the Mishnah, in a document whose entire purpose is to keep the chain of custody attached to the content. The rest of this page is what happens when you take that seriously and build the edge.

The chain of transmission

Teacher-to-student links are machine-extracted from the corpus — not curated by hand and not historically vetted — then normalised so that a single TEACHER_OF direction carries the whole chain. Between rabbi nodes there are 2,504 such links (1,191 TEACHER_OF plus 1,313 STUDENT_OF, reversed), joining 2,185 distinct sages. Those links do not form one tidy tree: the largest connected component holds 1,163 sages, 53.2% of everyone in the chain. The rest sit in smaller islands — which is itself the honest finding.

The network below is that largest component, ranked by transmission degree, top 150 taken, with all 288 teacher→student edges among them. Drag a node; hover for a name; scroll to zoom.

Verified chains present in the export

Hillel does not appear as a node in this 150-node export — a direct consequence of the transliteration problem described below rather than an absence in the tradition. The export file itself is public: /assets/data/transmission.json (150 nodes, 288 links, 28.5 KB).

What is in the graph

Counting labels in a multi-labelled graph is a trap — most nodes here carry several labels at once, so the usual labels(n)[0] shortcut silently undercounts. These figures come from a per-label membership query instead:

And the relationships that do the work — citation, thematic linkage, commentary, and transmission:

Retrieval is ranked, not arbitrary: PageRank has been computed across the graph's knowledge-base layer and 66,078 KBNodes — 100% of them — carry a score, so a query can be answered with the most structurally central sources rather than whatever matched first.

Maintained by surgery, not accretion

A graph like this degrades unless it is actively repaired. In July 2026 the teaching layer went through a measured overhaul, staged on a non-live copy first and cut over only after verification: the share of teachings orphaned from the rest of the graph fell from 76.96% to 4.60% — 12,408 teachings wired in — with 28,791 new teaching edges added and zero teaching nodes lost. The before-and-after report is preserved as a documented artifact, and the discipline it followed — never operate on the live instance, measure both sides, keep the rollback path — is the same revision-without-destruction principle the rest of this site describes.

Honest limits

A graph built by extraction is a graph with extraction's problems in it. These are the ones this dataset actually has, and they are the reason the headline numbers above lean on the clean, fully-covered fields rather than the noisy ones.

The safest single sentence about this dataset — every figure in it verified, none of it dependent on noisy labels — is this: 17,147 rabbinic teachings, each carrying its Hebrew source, a citable Sefaria reference, and a vector embedding — linked to 33,323 verses inside a 157,143-node graph.

Sources and attribution

Text and reference data are drawn from Sefaria and used under the Creative Commons Attribution 4.0 license. Every teaching node stores its own license, source and canonical Sefaria url, so attribution travels with the data rather than living only on this page. Sefaria's work in making the Jewish canon openly available is what makes a graph like this possible at all.

Corrections and questions: joshua@digitaltwinpro.com. If a number here disagrees with the database, the database wins and this page gets fixed. See the full research and data estate for everything else that sits behind it.