Google OKF ships three sample bundles built from BigQuery public datasets. Hands-on guide to ga4_obfuscated_sample_ecommerce and bitcoin_blockchain— sample queries, limitations, and how OKF turns raw tables into agent knowledge.
Open Knowledge FormatBigQueryGoogle AnalyticsBitcoinPublic DatasetsAI Agents
When Google launched Open Knowledge Format (OKF) v0.1 in June 2026, it did not ship an empty spec—it included three living sample bundles produced by a reference BigQuery enrichment agent:
This guide goes hands-on on the two datasets you linked: GA4 e-commerce and Bitcoin in BigQuery—what they contain, how to query them, and how OKF turns raw tables into agent-readable knowledge graphs.
ga4_obfuscated_sample_ecommerce — Nov 2020–Jan 2021
Bitcoin dataset
bitcoin_blockchain — full history, updates ~every 10 min
GA4 tables
events_* (sharded by date)
Bitcoin tables
blocks_raw, transactions
Cost
BigQuery Sandbox / free tier sufficient for samples
OKF value
Pre-linked concept pages vs raw schema discovery
Why Public Datasets Make Good OKF Demos
BigQuery public datasets are ideal OKF teaching material because they are:
Real-world shape — schemas, joins, and metrics agents actually need
Free to query — no data ingestion required
Well documented — Google publishes schemas and sample SQL
Diverse domains — web analytics (GA4) vs immutable ledger (Bitcoin)
OKF's bet: an agent answering "How do we compute weekly active users?" or "What columns link transactions to blocks?" should read curated concept pages with cross-links—not rediscover schema from INFORMATION_SCHEMA every session. That is the Karpathy LLM Wiki compile-once pattern at organizational scale.
SELECTCOUNT(*) AS event_count,
COUNT(DISTINCT user_pseudo_id) AS user_count,
COUNT(DISTINCT event_date) AS day_count
FROM `bigquery-public-data.ga4_obfuscated_sample_ecommerce.events_*`
BigQuery shows bytes processed before you run—useful for cost estimation. Valid queries display a check mark.
What agents learn from the OKF bundle
The GA4 OKF bundle (in knowledge-catalog) typically maps concepts like:
OKF concept type
Examples
Dataset
ga4_obfuscated_sample_ecommerce overview
Table
events_* schema, sharding pattern
Metric
Sessions, purchases, user_pseudo_id uniqueness
Event
purchase, add_to_cart, page_view parameters
Each file carries YAML frontmatter (type, title, description, resource, tags) and markdown cross-links—e.g., from a purchase event page to the ecommerce parameters page.
Since February 2018, Google has published the full Bitcoin blockchain to BigQuery for analytics. Allen Day and Colin Bookman announced on the Google Cloud Blog:
The Bitcoin blockchain data are now available for exploration with BigQuery. All historical data are in the bigquery-public-data:bitcoin_blockchain dataset, which updates every 10 minutes.
Property
Value
Project path
bigquery-public-data.bitcoin_blockchain
Tables
blocks_raw, transactions
Update cadence
~every 10 minutes as new blocks broadcast
Also on
Kaggle (BigQuery Python client in Kernels)
Why BigQuery for blockchain?
Bitcoin is an immutable distributed ledger with strong OLTP properties (atomic transactions, durability) but weak OLAP for ad-hoc reporting on money flows. BigQuery adds:
Real-time extraction from the ledger
Denormalized storage for easier exploration
Data Studio / Looker visualizations
OKF bundles capture this OLTP vs OLAP distinction so agents understand why the dataset exists—not just table names.
Network fundamentals queries
Google's original post highlighted network valuation analytics:
Analysis
Insight
BTC transacted per day
Economic activity on-network
Recipient addresses per day
User growth proxy (Metcalfe's Law)
NVT Ratio
Network Value to Transactions—valuation metric
Mining difficulty vs search volume
Fundamental + attention correlation
These become metric concept pages in the OKF bundle with formulas and caveats pre-documented.
Famous transaction: Bitcoin pizza (May 17, 2010)
Laszlo Hanyecz bought two pizzas for 10,000 BTC. The transaction is permanently recorded:
Transaction ID:a107...d48d (full hash in blockchain)
From:1XPT...rvH4 → To:17Sk...xFyQ
Google visualized input transfers up to 4 degrees before the pizza purchase—red circle for Hanyecz's address, blue for others, arrow width proportional to BTC flow. The OKF bundle can link entity pages (Hanyecz, pizza purchase event) to transaction schema pages.
Anomaly detection: duplicate transactions
One transaction appears in two blocks—impossible under current rules:
sql
#standardSQL
SELECT*FROM (
SELECT
transaction_id,
COUNT(transaction_id) AS dup_transaction_count
FROM `bigquery-public-data.bitcoin_blockchain.transactions`
GROUPBY transaction_id
)
WHERE dup_transaction_count >1
Why? Early Bitcoin used BerkeleyDB (non-unique keys). After Satoshi left, the team switched to LevelDB and implemented BIP-0030 to prevent duplicate transaction IDs. Legacy duplicate entries remain in historical data.
This is exactly the kind of contradiction / anomaly an OKF lint pass or concept page flags—agents get context without rediscovering Bitcoin history each query.
What agents learn from the OKF bundle
OKF concept type
Examples
Dataset
bitcoin_blockchain overview
Table
transactions, blocks_raw columns
Metric
NVT Ratio, daily transaction volume
Entity
Notable addresses, pizza purchase
Anomaly
Pre-BIP-0030 duplicate transaction_ids
Bundle 3: Stack Overflow (brief)
The third OKF sample maps the Stack Overflow public dataset on BigQuery—questions, answers, tags, votes. Same pattern: enrichment agent drafts concept docs for tables and common join paths (e.g., questions → answers → users). Browse the bundle in knowledge-catalog for the full file tree.
Raw BigQuery → OKF Bundle: The Pipeline
Google's reference BigQuery enrichment agent (in the repo) automates:
snippet
BigQuery dataset
↓ (walk tables/views)
Draft OKF concept .md per table/metric
↓ (second LLM pass)
Enrich with citations, joins, descriptions from authoritative docs
↓
Commit conformant OKF bundle → git
↓
Visualize with static HTML graph viewer
↓
Ingest into Cloud Knowledge Catalog for GCP agents
You can replicate this for your own datasets—the samples are templates, not products.
Upfront enrichment cost; reliable at moderate scale
For personal wikis under ~100K tokens, context often beats RAG. For org-wide data catalogs with hundreds of tables, OKF + Knowledge Catalog hybrid search scales further.
Browse OKF bundles — compare markdown concept pages to raw tables
Open HTML visualizer — see the knowledge graph
Point your agent — add bundle path to CLAUDE.md or project instructions
Summary
Google's OKF sample bundles are not synthetic toys—they map real BigQuery public datasets agents will encounter in the wild:
GA4 e-commerce — obfuscated Google Merchandise Store events (Nov 2020–Jan 2021), ideal for learning ecommerce event schemas
Bitcoin blockchain — full ledger history since 2018, ideal for OLAP-on-ledger patterns and anomaly documentation
Query the raw data to build intuition. Read the OKF bundles to see how compile-once knowledge beats per-query schema discovery. Then apply the same enrichment pipeline to your own BigQuery project.
