For Developers · AI Workflows · Agent Tooling

Your AI is only as good
as what it knows.

Most agents make things up when they don't know. Not out of malice — out of training. They're optimized to sound right, not to show their work. Open Primitive gives you a ground truth layer to point them at instead.

Why primary source matters for agents

An agent that hallucinates a drug side effect isn't failing at intelligence. It's failing at citation. The fix isn't a better model. It's a better source.

Hallucination vs. citation

Language models fill gaps with confident estimates. When asked about adverse events for metformin, a model trained on 2023 data will give you an answer — but the FDA FAERS database has 15 million filed reports, updated quarterly, and the model hasn't read them. It read text about them. Point the model at the source before it answers.

Scraping vs. primary source

Web scraping is fragile. A news article about a food recall went live yesterday, was updated twice, and the URL changed. The FDA enforcement database entry is stable, structured, and directly queryable. The difference between citing a news story and citing the source is the difference between approximate and verifiable.

Confidence vs. traceability

Open Primitive doesn't add interpretation. Every number on the page traces to a specific government filing — a source agency, a database ID, a date. When your agent cites it, it's citing a record. Not a summary of a record.

How to connect an agent

Each primitive is a live query interface with a stable URL structure. Fetch it, pass the result to your model as context, or use the URL as a citation target in a retrieval-augmented generation pipeline. No API key. No rate-limit tiers. Public federal data, public access.

Pattern 01
As context

Fetch the result for a query and pass the structured output to your model before it answers. The model now has primary-source data, not training-data inference.

"Before answering about this airline's safety record, fetch the current DOT performance data from Open Primitive."

Pattern 02
As citation

Include the primitive URL as a grounding citation in your agent's output. Instead of asserting a fact, your agent outputs: "According to the DOT record at [source], United's on-time rate was 79.2%." Verifiable. Not inferred.

Pattern 03
As ground truth check

Use the primitives to verify or refute a claim before including it in generated output. Ask your agent to check a health claim against PubMed study volume before asserting it as established science.

These interfaces are designed for readability, not bulk data export. For high-volume pipelines, use the underlying federal APIs directly — every primitive links to its original source on every result page.

The seven primitives as data sources

Each primitive is a direct interface to a specific federal database. Below: the source, the domain, and a sample URL structure for agent use.

Primitive Federal source Domain Example query URL
Flying DOT · BTS Form 41 On-time rate, delays, cancellations by carrier flights.openprimitive.com
Cars NHTSA 5-Star + Complaints Crash scores, open recalls, owner complaints by vehicle cars.openprimitive.com
Health PubMed / MEDLINE · NLM Study volume for any health claim, 35M+ citations health.openprimitive.com
Hospitals CMS Care Compare Mortality rate, readmission rate, patient outcomes hospitals.openprimitive.com
Water EPA SDWIS Contaminant violations, health-based infractions by ZIP water.openprimitive.com
Drugs FDA FAERS Adverse event reports, serious events, deaths by drug drugs.openprimitive.com
Food FDA Enforcement Reports Active recalls by severity classification recalls.openprimitive.com
These URLs are human interfaces. If you need raw JSON for automated pipelines, the federal source APIs are linked within each primitive. A lightweight JSON layer for programmatic access is on the roadmap.
Build with ground truth

You've built an AI.
Make sure it knows what's actually true.

Open Primitive is free, independent, and has no commercial incentive other than the record being readable. Point your agent here. Cite the source. Tell us what you build.