s0nar labs collects data from public sources, normalizes it into one schema and delivers it as structured feeds. Agents read current facts at run time, cite where each value came from and know how current it is.
Coverage includes the areas below. Refresh intervals vary by source, from 30 seconds to daily updates. Tell us which feeds, history and refresh intervals your project needs.
Spot prices, perpetual futures, cross-venue funding rates, lending and staking yields.
Market prices, resolutions and the largest moves.
Network gas, large wallet transfers, protocol TVL and ETF flows.
SEC Form 4 insider filings, the US Treasury yield curve and the US macro release calendar.
Industry headlines, the Hacker News front page, trending assets, App Store rankings and a market sentiment index.
US weather alerts and live NBA scores.
{
"source_id": "rates_us_treasury_curve",
"category": "macro rates",
"sym": "UST_10Y",
"ts_data": "2026-10-20T00:00:00+00:00",
"ts_extraction": "2026-10-21T13:02:11.418000+00:00",
"raw_payload": {
"tenor": "UST_10Y",
"rate_pct": 4.07,
"date": "2026-10-20"
},
"status": "ok"
}Illustrative record; not a live response. Here ts_data represents the observation date, not the exact publication time.
Records identify the source and collection time. Source or event time is included where available, with its meaning and precision defined for each feed. Market records add a SHA-256 integrity hash, and messages for on-chain consumers are signed with Ed25519.
A shared record structure gives your application a consistent way to read different sources. Source identifiers, timestamps and status fields let your team build one integration and handle updates as they arrive.
We do not interpolate, merge or fabricate values. When a source fails, its feed is marked stale with the time of the last good value, so an agent knows how current its data is.
A REST API with latest, history and batch endpoints. Batches fit a byte budget for model context windows and carry an as-of time and a staleness flag. Signed publishing serves on-chain consumers.
A model's trained knowledge does not automatically track changes in the world. Markets move, companies publish results and new events unfold. Without fresh input, a model can produce an outdated or incorrect answer that still sounds convincing.
AI systems now act: agents open tickets, write reports, monitor markets and send alerts based on what they read. An outdated value can lead to an unnecessary alert or a missed event. Fresh inputs and explicit timestamps help teams decide when an automated action is appropriate.
The practical answer is to separate what a model knows from what it reads. The model supplies reasoning and language. Current facts arrive at run time with source identifiers and timing information.
Large text, code and media corpora teach a model language and general knowledge. Smaller curated sets adapt it to a domain or a task. Adding new knowledge to the model's weights requires further training; retrieving new facts at run time is a separate process.
Before answering, the system searches a document store or knowledge base and passes the relevant passages to the model. Answers can cite their sources, and updating the store updates the answers without retraining.
Agents call APIs and tools during a task: a price, a filing, a forecast, a score. They need data that is current, structured and small enough to fit in the model's context window.
Test sets measure whether a model or an agent answers correctly. For questions about the present, the correct answer changes over time, so the test data has to record when each value was true.
Our data feeds serve retrieval and agents: they deliver current values at run time, with the timestamps an evaluation needs. Our document intelligence work serves retrieval over an organization's own documents. These services give teams a practical way to connect models to information that changes after training.
Every feed has a published refresh interval, from 30 seconds for market prices to once a day for the US Treasury yield curve. Each batch states when it was assembled and carries a flag that turns on when the feed falls behind its interval, so an agent can decide whether a value is current enough to use.
Records identify the source and collection time, with source or event time where available. The meaning and precision of those timestamps depend on the source; a daily observation date is not an exact publication time. Market records add a SHA-256 integrity hash, and messages published for on-chain consumers are signed with Ed25519.
A useful feed starts with a clear specification: the sources, fields, refresh interval, history and volume your application needs. Define how it will handle missing values, duplicates and delayed updates, and what a successful delivery looks like.
We work from that specification to define the schema, delivery method and acceptance checks. We agree on sample availability and validation steps before production delivery, so your team can plan the integration.
A data feed is a continuously updated stream of records from one or more sources, delivered in a fixed schema. For an AI system, the feed supplies external facts in a form an agent can read through an API, including updates beyond the model's training data.
Each feed has its own refresh interval, from 30 seconds to 24 hours depending on how often the source publishes. Every batch states when it was assembled and whether the feed is behind its interval.
Through a REST API with endpoints for the latest value, recent history and batches sized to a byte budget. Access is controlled with API keys. For on-chain consumers we also publish signed messages.
Yes. Send us the source, fields, refresh interval and history you need. We assess feasibility and propose a schema and delivery plan for your application.
Need a source we do not cover? We build feeds to a written specification: sources, fields, refresh interval and acceptance rules.
Request a sample feed