Products Document intelligence

Ask your documents a question and get an answer that cites the page, with nothing leaving your network.

Radiant is a self-hosted document intelligence platform. It reads your documents, finds the passages that matter and answers questions from them, citing the document, page and section behind every statement. Access follows each document's classification, and everything, including the language model, runs on your own GPU hardware with no route to the internet.

Request a demo See how it works

A Radiant overview answer listing the recurring problems across a collection of site reports, with every theme cited to the documents behind it

Until an answer starts to appear: About 1.3s
Cites the document and page it came from: Every answer
To open a network map of a million links: Under 0.5s
Internet connections needed: 0

Times are measured for one user on NVIDIA DGX Spark, the most compact system Radiant runs on. Server-grade hardware is faster still. Longer answers take longer, and an overview of a whole collection takes about 20 seconds.

Answers your analysts can check

Answers that show their working

Ask a question in plain language. Radiant finds the passages that answer it and writes a short answer from those passages alone, citing each one. Every citation is checked against the source text before the answer is shown, and when your documents do not contain the answer, Radiant says so rather than guessing.

Citations to the document, page and section, checked against the source
The cited region outlined on the page, scanned pages included
Follow-up questions in a conversation, with answers that stream as they are written
Overviews of a whole collection, with every theme cited

The Radiant document viewer with the cited passage outlined on a two-column technical specification

Each person sees only what they are cleared to see

Every document carries a classification, and can carry access groups and caveats such as UK EYES ONLY. Radiant enforces them on every path: search, answers, chat, collection overviews and the knowledge graph. A document someone may not see is simply not found.

Classification levels and caveats managed by your own administrators
A clearance, access groups and a role in each collection for every person
API keys for your own systems, each with a clearance of its own
Overviews written only from the documents the reader may see

Radiant administration: OFFICIAL, OFFICIAL-SENSITIVE and SECRET classification levels and a UK EYES ONLY caveat, with how widely each is used

See who is connected to whom

The network map draws out the people, organisations, places, phone numbers, e-mail addresses and accounts in a collection, from the exports your teams already have, such as call logs, messages, transactions and registers, and from the relations found in documents. Every link shows the evidence behind it, and each analyst sees only what their clearance allows.

Imports CSV, Excel, Parquet and JSON exports
Paths, communities and summaries whose every sentence cites its records
Different spellings and identifiers united without altering the original records
Opens in under half a second, even at a million links

The Radiant network map: people, phone numbers, places and organisations drawn as a graph, with a timeline above

From a pile of documents to a cited answer

Radiant does the reading, so your teams can ask questions instead of searching file by file.

PDF, Office, scans

Reads almost any document

PDFs, Word, PowerPoint, Excel, web pages and images are converted with their layout, tables and page numbers intact. Hard pages, such as scanned Arabic or Chinese, are read again by the system's own vision for accuracy.

Hybrid search

Finds the right passage

Search by meaning and by keyword are combined, then the best candidates are ranked again, so the passage that answers the question rises to the top.

Overviews

The whole collection at once

Broad questions, such as what risks recur across a set of reports, are answered from a summary of every document the reader may see, with each theme cited.

Deep answers

Joins facts across documents

For questions that span several documents, Radiant can also follow its knowledge graph of the people, places and things your documents mention.

Connectors

Stays up to date

Collections refresh on a schedule from a watched folder and from on-premises Confluence, SharePoint Server, MediaWiki, WebDAV, GitLab and Jira, and everything is also available through a secure API.

Encrypted export

Moves between sites

A whole collection can be exported as an encrypted file and imported at another site without being processed again.

Built for classified and sensitive material

Radiant is built for government, defence and regulated teams whose documents cannot leave the site. Its controls are mapped to the NCSC Cyber Assessment Framework, Secure by Design and the UK codes of practice.

Sign-in for every user, and an API key with its own clearance for every system
Encrypted connections through a single entry point
No container can send data off the host, confirmed by an automated check
Classification, group and caveat checks on every read path, failing closed
A tamper-evident audit trail of sign-ins, questions, uploads and changes
Question and document text are never written to the logs
All AI models from US and European developers, with their licences included
A signed installation package that needs no network at any point

Typical uses

Investigations and casework
Policy, regulation and case files
Technical manuals, specifications and maintenance records
Reports and archives in several languages

Runs on the hardware you choose

Start small with DGX Spark

Radiant is optimised for NVIDIA DGX Spark, a compact desk-side system that runs the whole platform, language model included, and is all a team needs to get started.
Scale up to the data centre

It runs just as well on DGX workstations and data-centre GPU servers such as the NVIDIA B300, provided there is enough GPU memory, and gets faster with every step up. We can help you choose.
Installed offline

One signed package carries the software, the AI models and an installer. The target machine needs no network at any point, during installation or after.

Questions buyers ask

What hardware does Radiant need?

Radiant is optimised for NVIDIA DGX Spark, a compact desk-side system, and is designed to scale up to larger DGX-class systems and data-centre GPU servers. Tell us about your environment, document volumes and number of users and we will confirm the right hardware.

Does Radiant need an internet connection?

No. It is installed from one signed package that carries everything it needs, and no part of the running system can send data off the host. The host's own network configuration stays under your control.

How does it stop people seeing documents they should not?

Every document carries a classification, and can carry access groups and caveats. Each person and each connected system has a clearance, and Radiant checks it on every path, from search and answers to overviews and the network map. A document someone may not see is reported as not found. Radiant holds no formal accreditation, so your own assurance process applies.

Can we trust the answers?

Every answer cites the passages it rests on, and each citation is checked against the source text before the answer is shown, so any statement can be confirmed in a click. When your documents do not contain the answer, Radiant says so rather than guessing. As with any system, check important details against the cited sources.

Which file types does it read?

PDFs, both born-digital and scanned, Word, PowerPoint, Excel, HTML, Markdown, CSV and images. Scanned pages are read with OCR, and hard pages, such as scanned Arabic or Chinese, are read again with the system's own vision.

Does it work across languages?

Yes. Ask in English and get cited answers from documents written in Arabic, Chinese or Spanish. It works best where those documents are not competing with many similar English ones, and we are happy to show you on your own material.

Where do its AI models come from?

All of Radiant's AI models come from US or European developers, including the models they were built on, and their licences ship with the product.

Can it connect to our existing systems?

Yes. Connectors keep collections up to date from watched folders and from on-premises Confluence, SharePoint Server, MediaWiki, WebDAV, GitLab wikis and Jira, and everything in the interface is also available through a secure API.

How is Radiant licensed and supported?

Licensing and support are tailored to each deployment, from a single team to an organisation-wide service. Talk to us about your requirements.

Can we try it before committing?

Yes. We can arrange a demonstration, or a pilot on your own hardware using your own documents.

See Radiant on your own documents

Talk to us about a demonstration, a pilot on your hardware, or help choosing the right hardware for your workload.

Request a demo

Also from Devsultants MetaVox: offline speech recognition. Speech to text in 1,600+ languages, without your audio ever leaving the building. Also from Devsultants MetaLex: private translation. Translate text and documents on your own hardware, with nothing leaving the machine. Also from Devsultants Cold Harbour: software supply chain security. Every third-party component verified, scanned and approved before your builds can use it.