Olo Commons

An open, verifiable knowledge base that an AI can think with, and that belongs to everyone.

Olo Commons grows every 15 minutes from sources the world already trusts: Wikipedia in English and Spanish, Wikidata, Wikinews, arXiv, PubMed, OpenAlex, Project Gutenberg, the Internet Archive (items under an open licence only) and Data.gov are live; connectors for Europeana, DOAJ, Open Library, Europe PMC and Wikimedia Commons are written and tested and start disabled until their SQL is run (docs/TRUST.md lists every source with its licence). Every unit of knowledge (an "atom") keeps its provenance: source, URL, licence, authors, publication date, a SHA-256 fingerprint of its content and the time it was retrieved. Atoms are embedded and linked to their nearest neighbours, so the Commons is a graph that thickens as it learns. Anyone can read it through one free API. Only allow-listed sources, fetched by the server, can write to it.

Live: https://heliusnetwork.app/olo

Since October 2026 every atom also carries trust and time: a status (candidate, verified, disputed, superseded, retracted), a computed credibility, validity dates, the revision of its source and what replaced it (docs/TRUST.md). Sources are re-read and nothing is overwritten. Sentence flags from Olo's apps, reviewed by a person, teach Zeus every day (docs/FEEDBACK.md); Zeus's measured numbers are in docs/ZEUS_SCORECARD.md. Olo, Zeus and the Commons together are Helius Network's WEBOCTA system (docs/WEBOCTA.md).

Why

Language models are good at reasoning and bad at remembering. The fix is not a bigger model; it is a shared, citable memory that is open to inspect, open to correct and open to build on, in the spirit of the Internet Archive and Wikipedia. Olo Commons is that memory for Olo (helius-olo.io and the Olo inside the Helius Network app), and for anyone else.

What is in this repository

• sql/001_olo_commons.sql: the schema for Postgres with pgvector. Sources registry, atoms, links, queue, run log, daily snapshots, projects, read-only API functions, Row Level Security, scheduled jobs.
• sql/002_trust_time_provenance.sql: status, credibility, validity dates, source revisions, supersession, the feedback table, retractions and the per-address cap on curiosity gaps. Run after 001.
• sql/003 to sql/017: the Olo brain, later connectors' registry rows, deduplication, and the cleanup files that apply the ingester's quality rules to atoms stored before each rule existed. Numbered, safe to rerun, run in order; each file's header says what it does and why.
• functions/olo-ingest: the ingester (Supabase Edge Function, Deno), run every 15 minutes in small batches. One discovery routine per source, using that source's own public API; a quality gate (text.ts) refuses stubs, catalogue records and pointers, and a licence check (licence.ts) refuses anything that is not public domain, CC0, CC BY or CC BY-SA where a source licenses per item. Embeds with gte-small inside Supabase; no external AI key.
• functions/olo-query: the read API. Meaning search merged with keyword search, provenance on every result, rate limiting by hashed address, CORS for known origins.
• functions/olo-answer: the answer API. Retrieves like olo-query, drafts two to four cited sentences with a hosted open-weight model, and has every sentence checked (supported, contradicted, not covered, with quoted evidence) by Zeus when a served adapter is configured, otherwise by the drafting model under the Zeus prompt, and says which. Uncovered questions become gaps the ingester fills next.
• functions/olo-feedback: sentence flags from the apps, bounded and rate limited, reviewed before they teach Zeus.
• functions/olo-brain-ask: questions people choose to share, scrubbed of personal details, shown as orbs in the public Olo brain (docs/BRAIN.md); unanswered ones feed the gap list.
• docs/API.md: how to call it. docs/TRUST.md, docs/FEEDBACK.md, docs/BRAIN.md, docs/ZEUS_SCORECARD.md, docs/WEBOCTA.md.
• SECURITY.md: what is enforced and how to report a problem.

Run your own

1. Create a Supabase project. In the SQL editor run sql/001_olo_commons.sql. Store your service-role key in Vault as service_role_key (the schedule uses it).
2. supabase functions deploy olo-ingest --no-verify-jwt and supabase functions deploy olo-query --no-verify-jwt. (This repository also ships .github/workflows/deploy-functions.yml, which deploys both after a human approves the run; see the comments at the top of that file for the two-step setup.)
3. Optional secrets: OLO_API_KEYS (comma-separated partner keys for a higher rate limit), OLO_SALT (for address hashing). For answers: OLO_LLM_API_KEY (an OpenAI-compatible provider key; without it olo-answer returns 503), OLO_LLM_BASE_URL (default Groq), OLO_LLM_MODEL (default openai/gpt-oss-120b), and OLO_ZEUS_URL once a Zeus adapter is served.
4. Trigger the first ingest by hand: select public.run_olo_ingest();. From then on it runs every 15 minutes (3, 18, 33 and 48 minutes past the hour).

Zeus: the verification model

zeus/ holds the open build of Zeus, Olo's claim verification layer: a rank-8 LoRA on Mistral 7B Instruct v0.2 (Apache 2.0) trained on examples made from the Commons, open fact-checking corpora, hand-written gold and reviewed feedback. It trains in the cloud every day when there is something new to learn (.github/workflows/zeus-train.yml, zeus/cloud/), is measured on a frozen split, on held-out VitaminC and on an LLM-AggreFact sample, and serves on Cloudflare Workers AI (zeus/cloudflare/). zeus/STATUS.md records each build; docs/ZEUS_SCORECARD.md is the current scorecard. Hand-written examples in zeus/data/gold/ are the most useful contribution you can make.

Build on it

• examples/cited-chat.mjs: a chatbot that asks the Commons first and cites.
• examples/notebook.py: a topic's atoms as a pandas DataFrame with provenance.
• examples/graph-export.mjs: the whole graph as GraphML or JSON.
• docs/API.md: the read API; docs/ARCHITECTURE.md: how it works inside.

Starting points for contributors

Issues labelled good first issue are small and real: a new source connector, a better clean(), a language other than English, tests for a connector. Templates: propose a source, report an atom, bug.

Adding a source

See docs/ADDING_A_SOURCE.md. In short: open a pull request that adds a discover routine in functions/olo-ingest/index.ts and a row in olo_sources. Requirements: a public API or dump, an open licence that allows redistribution with attribution, no personal data, and a trust value with a one-line justification. A source that is not enabled in the registry is never fetched.

Licences

• Code: Apache License 2.0 (see LICENSE).
• Data: each atom carries the licence of its source and must be used under that licence. The Commons as a compilation is offered under CC BY-SA 4.0.

Governance

Sources, trust levels and project approvals are reviewed by Helius Network Ltd for now, in public, through issues in this repository. The intent is to move that to an open editorial group as contributors arrive.