Jew.tech → Data
Research & data estate
This is the corpus the rest of the site stands on, counted rather than estimated. Every figure below came from a query whose exact text is kept in a claims ledger with its result and its verification date; all of them were re-run on 2 August 2026. Where a number would flatter the work if stated loosely — deduplicated documents, indexed vector search, live versus stale copies — the looser version is not published and the reason is given instead.
Online right now, and what is not
An estate description that quietly counts unplugged drives is a fiction. Here is the split, as of 2 August 2026:
| Holding | State | Where |
|---|---|---|
| Torah knowledge graph (Neo4j) | Online — queried live for this page | studio1 |
| File + provenance catalog (PostgreSQL 17 / pgvector, 38 GB) | Online | studio1 |
| Web inventory (PostgreSQL) | Online | studio1 |
| Vector store (Qdrant, 2 collections) | Online | studio1 |
| Document stores (ArangoDB, 4 databases) | Online | studio1 |
Research corpus (SQLite, research.db) | Online | local working copy |
| Publisher PDF harvest + model backups (Synology NAS) | Offline right now — not counted in the totals above | NAS |
| Graph replicas on two Mac minis | Offline right now — not counted in the totals above | mini1, mini2 |
The offline holdings are real and are named here for completeness, but no count from them appears in any total on this page. They will get their own verified rows in the ledger before any number from them is published.
The Torah graph
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. That is the whole claim, and every part of it is at 100% coverage. Around it sit 215,426 relationships across a 233-label schema, 15,910 verse citations, and PageRank scored over all 66,078 knowledge-base nodes. The text comes from CC-BY Sefaria sources.
The full breakdown, the sample node, the interactive transmission network and the extraction caveats live on its own page: the Torah knowledge graph.
Research corpus
A literature pipeline that ingests papers, extracts concepts and ideas from them, and assembles the result into a graph. 22,841 papers deduplicated on ingest produce 99,620 concepts, 81,966 extracted research ideas, and a graph of 202,816 nodes and 336,609 edges. 1,595 papers are rated high-relevance.
| Measure | Count |
|---|---|
| Papers ingested (deduplicated) | 22,841 |
| Papers rated high-relevance | 1,595 |
| Concepts extracted | 99,620 |
| Research ideas extracted | 81,966 |
| Paper–concept links | 169,549 |
| Graph nodes | 202,816 |
| Graph edges | 336,609 |
One database is the write target: research.db. The graph copies in Neo4j and
ArangoDB are downstream exports, which is why the numbers in those stores lag the ones
above.
File and provenance catalog
34,457,680 files indexed across five machines, with a provenance layer on top that tracks which physical copy a page came from and which OCR run produced its text. 234,591 of those files are PDFs, totalling 10,232,735 pages.
| Provenance layer | Count | What it tracks |
|---|---|---|
| Page-link relations | 42,669,045 | Page-to-page and page-to-source links. |
| Pages with provenance | 10,233,053 | Pages carrying a traceable origin record. |
| OCR runs | 274,098 | Each text extraction, kept as its own record rather than overwritten. |
| Physical copies tracked | 127,234 | Distinct scans/editions of a work. |
| Distinct works | 43,174 | The work-level identity above the copies. |
| Book chapters | 2,445 | Chapter-level structure inside catalogued books. |
| Machine | Files catalogued |
|---|---|
| macbookpro | 23,102,817 |
| JoshuascStudio1 | 11,261,114 |
| JoshuascStudio3 | 34,674 |
| mini1 | 32,290 |
| Joshuas-Mini-2 | 26,785 |
| Total | 34,457,680 |
count(DISTINCT sha256) returns zero — so 234,591 is a
count of PDF files catalogued, and the same document appearing on two machines is
counted twice. Deduplication is future work; until it is done, the word "unique" does not
appear next to any of these figures.
Vector stores
Two Qdrant collections holding 34,865 vectors between them, at two different embedding dimensions because they serve two different purposes: page-level retrieval over OCR'd material, and book-level retrieval over the corpus.
| Collection | Points | Dimensions |
|---|---|---|
midrash_pages | 20,933 | 384 |
corpus_books | 13,932 | 768 |
| Total | 34,865 | — |
indexed_vectors_count: 0 — their HNSW indexes are not built, because each
segment sits below the 10,000-vector threshold at which Qdrant builds one. Queries are
brute-force and correct; at this scale that is fine. No claim of "HNSW-indexed vector
search" is made.
Document stores
ArangoDB holds the document-shaped material: OCR output, extracted teachings, layout analysis, and an entity/relation graph over the Kroll corpus.
| Database | Collection | Documents |
|---|---|---|
| corpus | ocr | 17,161 |
| corpus | extracted_teachings | 2,245 |
| corpus | ocr_layout | 2,345 |
| vault | vaultteaching | 13,397 (stale snapshot — live figure is 17,147) |
| research | papers | 8,229 (stale snapshot — live figure is 22,841) |
| kroll | relations | 4,653 |
| kroll | entities | 3,143 |
The two snapshots marked stale are exactly that: older exports that have not been refreshed from their write targets. They are listed rather than hidden, and their live counterparts are named beside them so no one mistakes a snapshot for the current state.
Web estate
90 web properties tracked, of which 76 are live and a further 5 are deployed on Workers — 82 live or deployed, 8 dead, 1 asleep for maintenance. 470 uptime probes and 383 DNS zones are configured across them, with 84 deploy-audit records and 76 theme-QA records on file, and 15 broken links currently outstanding and unfixed.
The Judaism, Torah and AI properties among them include jew.tech, torahtech.com, torahchain.com, torahtwin.com, talmudiclogic.com, gematriagraph.com, openrabbi.com, hashemlibrary.com, alephinity.com and tikkunology.org.
How this is served
All of the online holdings above run self-hosted on one machine — a Mac Studio reachable only over a private network — with no managed cloud database anywhere in the stack. Eleven containerised database engines do the serving: six Neo4j instances (the Torah graph, its earlier snapshot, and four smaller graphs), three PostgreSQL 17 instances with pgvector (the file, provenance, and web-inventory catalogs), Qdrant for vectors, and ArangoDB for documents. Four of the eleven were queried directly to produce this page; the container inventory behind the count is in the fleet map.
That arrangement is the point rather than a constraint. A corpus you own, on hardware you control, can be counted exactly, re-counted on demand, and corrected in public — which is the only basis on which any of the numbers here are worth stating.
How these numbers are made
Every figure on this page has a row in a claims ledger recording the claim, the exact query that produced it, the result, and the date it was verified. The ledger is re-run before publication; where a re-run disagrees with the page, the page is wrong and gets fixed. Three standing rules govern what appears here:
- A number with no ledger row is not published, however plausible it is.
- A caveat found during verification is published beside the number, not buried — files rather than unique documents, exact rather than indexed search, snapshot rather than live.
- Holdings that are offline are labeled offline and excluded from every total.
All figures verified 2 August 2026. Corrections and questions: joshua@digitaltwinpro.com.