How I Built a Visual Knowledge Graph Encyclopedia (1,318 Entities, 12,146 Cited Claims) Without a Traditional Database
When I started building ATLAS, I wanted to fix a frustrating pattern when studying history, politics, and economics online: you read an article on Marxism, Liberalism, or the Vargas Era, jump to a scattered summary elsew
When I started building ATLAS, I wanted to fix a frustrating pattern when studying history, politics, and economics online: you read an article on Marxism, Liberalism, or the Vargas Era, jump to a scattered summary elsewhere, lose track of who influenced whom, and have no quick way to tell whether a claim is a documented fact, an attributed thesis, or a contested interpretation.
The default architecture would be Postgres + a Node/Go API + an admin panel. For an open-source project maintained by one personβand designed to stay fast and cheap without a 24/7 backendβa traditional database adds more operational drag than value.
Instead, I went with content-addressed static releases on Cloudflare R2 (SHA-256 hashed JSON bundles), served to an Angular 21 front-end with 1,413 prerendered routes on Cloudflare Pages.
Real numbers from the current release (v0.7.0)
- 1,318 typed entities (people, works, concepts, parties, institutions, events, countries, places)
- 12,146 claims, every single one with a citation β split into 8,278 documented facts, 1,544 contested claims, 1,171 attributed theses, and 1,153 interpretations
- 1,482 sources
- 3,003 typed relations in the graph
- 14 continuous readings and 21 guided itineraries
1. Content-addressed releases on R2 (no DB, no broken cache)
Instead of querying an API on every click, the build pipeline (npm run content:build) validates all 24 collections with Zod, resolves graph edges, splits the data into specialized bundles (readingBundle, networkBundle, and the full bundle), and hashes every artifact with SHA-256.
The Git repo only stores a tiny pointer file (content/content-pointer.json):
- Every production build reads the exact release hash pinned in Git.
- Because artifact filenames contain their SHA-256 hash, edge caching can be aggressive and immutable without risking version skew.
- The 1,078 Wikimedia images referenced by entries are also mirrored on R2 by content hash (
m/<sha256>.jpg), avoiding broken external hotlinks.
The trade-off I'm still working on (and tracking as an open issue): readings currently fetch a heavier bundle than they strictly need for a single page. Next step is finer route-level splitting.
2. Local LLMs for mining draftsβnever as autonomous authors
A common question is whether ATLAS content is AI-generated. It isn't.
I run local models via Ollama (qwen3 / gemma2) plus GPU embeddings (BAAI/bge-m3) strictly in offline ingestion scripts (scripts/ingestion/), behind three guardrails:
-
Closed corpus: models only see primary texts and curated sources extracted locally (constitutions, primary documents, Wikipedia CC BY-SA 4.0, SEP/IEP) via
firecrawl-local. -
Mandatory verbatim quotes: every proposed claim or relation must include a
quotestring. A deterministic validator checks that the quote appears verbatim in the source passage and explicitly names both endpointsβotherwise it is rejected automatically. -
Human review gate: accepted drafts enter as
pending(ok: null) and never publish automatically.
3. Four views over one graph
Because all entities and edges live in one typed model, the same graph drives four different ways to explore:
-
Readings (
/leituras): long-form continuous prose with sources and claim types inline. -
Itineraries (
/itinerarios): question-driven step-by-step routes through the graph. -
Network Graph (
/grafos): interactive network with path finding between any two entries. -
World Map (
/mapa): canvas globe with historical connections and a 1750β2026 timeline.
Want to explore or contribute?
Code is MIT, content is CC BY-SA 4.0:
- Live: atlas.bellussi.dev
- Repository: github.com/bellussi1/ATLAS
- Good first issues: browse open issues (includes both code/perf tasks and non-code source/image audits)
If you've shipped large static datasets in Angular or React SPAs, how do you split your data bundles to keep initial load fast without spinning up a query API?
Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.