Registry data, queried in plain language.
Regis takes a question in natural language, retrieves from primary government registry records across 200+ countries, and returns the answer with the filing behind every value.
- ✓ResolvedGlobal Data Intelligence Limited → GB 09410808
- ✓RegisterCompanies House, United Kingdom
- ✓Readingstatus, company type, incorporation date, registered office
The model can only return what it retrieved.
A general-purpose model answers a company question from whatever it absorbed in training. There is no filing behind that answer, no date attached to it, and no way for a reviewer to check it. Regis is built the other way round: the question triggers retrieval against registry records, and the answer is assembled from what came back.
Retrieval first
No answer without a record. The question is resolved to entities and fields, those fields are read from primary registry records, and the answer is written from the retrieved values. Nothing in the response originates in model weights.
Refusal is a result
Empty is a valid response. Where the register does not publish a field, Regis returns the gap rather than an estimate. An empty answer with a stated reason is safer in a case file than a confident one you cannot source.
Same question, same answer
Reproducible for review. Answers are reconstructed from the underlying record, so a reviewer running the query again gets the same values, with the retrieval timestamp attached to show what the register held at the time.
Structured output
JSON, not paragraphs. Alongside the written answer the response carries the resolved entities, the fields used and the source records, so it can be written straight into a case management system without parsing prose.
Reaching the same records through a general assistant.
Company data can be pulled into a general-purpose model through a connector. It works. What it changes is how many suppliers sit between the question and the register, what the retrieval is billed as, and how much of the answer a reviewer can trace.
| Regis API | Data provider plus a general assistant | |
|---|---|---|
| Source of the answer | Registry records retrieved at request time, with no company facts held in weights | Records passed into a model context, assembled by a model that also answers from training |
| Citation level | Field level. Each value carries its register, reference and retrieval time | Document level, where the system cites at all |
| Missing field | Returns null and names the register that is silent | Depends on the assistant's behaviour |
| Commercial contracts | ✓One | Two |
| Data processing agreements | ✓One | Two |
| Assistant seat licences | ✓None | Required |
| Pricing basis | ✓Per result | Per token consumed |
| Bulk retrieval | ✓Direct, no inference | Through the model context |
| Data path | ✓Single supplier | Crosses a third party |
| Vendor review | One vendor on the register | Two vendors to assess, two answers on where the data travelled |
Both columns return the same registry records.
Ask a general-purpose AI assistant for five years of financials on one company. Its accounts carry about 250 separate figures a year, so 1,250 figures have to travel into the model’s context window before it can say anything. That is roughly 19,000 tokens, billed at inference rates, to move data that was already structured.
0
FiguresFive years of accounts for one company0
TokensWhat those figures cost to put in front of the model0
TokensThe same request across a thousand companies- Every figure is pulled into the model’s context before it reaches you — the retrieval itself is priced as reasoning
- Ask a follow-up and the same context is sent again and billed again
- On top of the token charge sits the assistant subscription, plus a seat for everyone who uses it
- A second commercial contract and a second data processing agreement
- The same figures come straight out of the register as structured data, with no model in the middle
- No tokens involved
- Charged per result rather than per token, so a thousand-company pull costs what moving data costs
- No subscription, no seats — one contract, one agreement
One dataset, two access patterns.
Global Database’s registry data is reachable two ways on the same API key. The structured endpoints take a company identifier and a named field, and return a fixed response shape. Regis takes a question and works out for itself which registers and fields answer it. Both read the same records.
Structured endpoints
You name the company and the field. KYB, officers, shareholders, group structure, financials, credit and bank verification each have their own path and a fixed response shape.
- Input
- Identifier or filters
- Output
- Fixed JSON schema
- Provenance
- include_provenance
- Best for
- Pipelines, batch runs
Regis query
You ask the question. Regis resolves the entities, decides which registers and fields answer it, retrieves them, and returns the answer with its sources. Multi-step questions resolve in one call.
- Input
- prompt as a string
- Output
- Answer plus sources
- Provenance
- Attached by default
- Best for
- Analysts, agents, apps
One POST, a token header, a string.
Authentication is a token in the header. There is no SDK to install and no query language to learn. If your service can make an HTTP request, it can call Regis.
# Ask a question. Get the answer and its sources. curl -X POST 'https://api.globaldatabase.com/v1/regis/query' \ -H 'Authorization: Token YOUR_API_KEY' \ -H 'Content-Type: application/json' \ -d '{ "prompt": "Is Global Data Intelligence Limited still active, and what is its registered office?" }'
prompt and returns the answer with its sources. The rest of the v2 API stays available for structured calls against the same records.What arrives with the answer.
The written answer is the readable part. The part that survives an audit is what sits underneath it.
{
"field": "status",
"value": "Active",
"source": {
"register": "Companies House",1
"jurisdiction": "GB",
"filing_reference": "CS01",2
"filed_on": "2026-01-28",
"retrieved_at": "2026-08-04T05:43:17Z"3
},
"resolved_entity": {
"name": "Global Data Intelligence Limited",
"registration_number": "09410808"4
},
"unavailable_fields": []5
}- 1source.register
Names the government body that publishes the record, so the answer resolves to the statutory source rather than to a vendor's own dataset.
- 2source.filing_reference
The document identifier as the register issued it. A reviewer can pull the original filing from the register without going through us.
- 3source.retrieved_at
Fixes what the register held at the moment the decision was made, which is the question asked months later when the record has moved on.
- 4resolved_entity.registration_number
Shows which of several similarly named companies the answer is actually about. Most disputes about company data are resolution errors, not value errors.
- 5unavailable_fields
Lists what the register does not publish for that jurisdiction, so a missing value is distinguishable from a value of zero or none.
include_provenance: true on enrichment, autocomplete and financials calls to get source metadata alongside the values. Documented in the v2 reference.Four hundred registers do not agree with each other.
Plain-language querying is the surface. What makes it answerable is the part that took years: reconciling registers that publish different fields, in different languages, on different schedules, under different definitions of the same word.
Five incompatible ways of naming the same class of object. None of them was designed to interoperate with the others.
Entity resolution
The same company appears under a trading name in one register, a legal name in another and a transliteration in a third. Resolving a question to a registration number is the step that decides whether the rest of the answer is about the right company.
Field definitions
Turnover is not the same line in every accounting standard, and a director is not the same office in every company law. Answers are normalised to a common dictionary while the original value stays attached.
Document extraction
A large share of registry filings are documents rather than data. Financial statements, shareholder registers and charge filings are read out of the filed document and structured into fields.
Update cadence
Registers publish on their own schedules, from continuous to annual. Records are synced daily, and the retrieval timestamp tells you which side of a register's update you are on.
Ownership traversal
A beneficial owner three jurisdictions up is a chain of registry lookups, not a field. Regis walks the chain and returns each hop with the register it was read from.
The questions your second line will ask.
Every one of these arrives before a data source is approved. The answers are here so you can send them on without waiting for us.
Where does the answer come from?
Primary government registry records, retrieved when the request runs. Not model recall, and not a score derived from someone else’s dataset.
Can we reproduce it months later?
Yes. Every value returns with its register, filing reference and retrieval timestamp, so the answer can be rebuilt against the original filing long after it was generated.
What happens when it does not know?
It returns null and names the register that does not publish the field. No estimate, modelled figure or comparable-company value is substituted for a missing filing.
Is our query data used for training?
No. Questions and payloads sent to the API are excluded from model training, with the controls set out in the data processing terms.
What certifications do you hold?
ISO 27001, and registration with the UK Information Commissioner’s Office. Security documentation and the data processing agreement are available under NDA before contract.
Can we redistribute the output?
Under a redistribution licence. Internal use, customer-facing display and redistribution are licensed separately, so contract scope is explicit rather than assumed.
Ask it something you already know the answer to.
That is the fastest way to judge it. Pick a company in your book, ask Regis a question you can verify, and check the filing it returns.
Common questions.
What exactly is Regis?
How is this different from asking a general-purpose model?
What happens when the register does not publish the field?
Which countries are covered?
Can I use it inside my own product?
Does it work with an AI agent?
How current is the data?
How do I authenticate?
Authorization: Token YOUR_API_KEY. Keys are issued from your account profile.