Elevating information to knowledge
ELYVRA turns scattered sources into structured knowledge and answers questions by citing the passage it used, with its document, version and validity.
Pre-launch: available on request for teams who want to try it.
Example · fictitious content
AI-generated answer
Question · How long does the product warranty last?
The warranty lasts twenty-four months from delivery. For wear parts the period is reduced to twelve months.
“The warranty lasts twenty-four months from the date of delivery.”
01 · The starting point
Documents, websites, mailboxes, archives and datasets: they are produced and received all the time, in different formats and systems.
The file names in the scene are examples.
01 · The starting point
Different formats, different rules, copies that contradict each other. Where does a fact come from? Is it still in force? Which passage does it refer to?
01 · The starting point
Structuring means extracting the content, giving it context and keeping track of it. Then the same body of knowledge serves search, answers and applications, without starting over every time.
Step 1 / 5
Collect from many sources
Sources arrive through uploads or scheduled connectors. Before processing, files are scanned for malware.
Step 2 / 5
Extract, normalize, classify
Text is extracted from the original format and put in order: headings, paragraphs, tables, attachments. Then the metadata is derived.
Step 3 / 5
Prepare for retrieval
Text becomes chunks of a useful size. Each one is indexed both by words and by meaning.
Step 4 / 5
Provenance, versions, time
Every chunk remembers where it comes from. Versions are never overwritten, and validity over time is data, not a guess.
Step 5 / 5
Make knowledge usable
With authorization already checked, knowledge is read by search, assistants and applications.
02 · The result
A single layer that remembers where every piece of information comes from, which version counts and since when. People, applications and AI all work on the same foundation.
ELYVRA builds this layer: with customers isolated from each other, independent of the type of data, with every domain described by a profile instead of written into the code.
Go to the platform mapThe Knowledge Fabric
Each item is a platform capability, shown with its status. Select an item to see its role. ELYVRA is in pre-launch: the implemented features are available on request for teams who want to try it.
01 · Sources
6 in the platform
02 · Processing
5 in the platform
03 · Knowledge
5 in the platform · 1 on the roadmap
Knowledge layer
Chunks with provenance
document · version · section
Every chunk knows which document, version and section it comes from: that is what makes verifiable citations possible.
04 · Retrieval
4 in the platform
05 · Applications
4 in the platform · 1 in development · 2 on the roadmap
Knowledge layer
Chunks with provenance
document · version · section
Every chunk knows which document, version and section it comes from: that is what makes verifiable citations possible.
Study material and quizzes from every knowledge base, linked to the original documents. Design complete.
Explore LearningNo dates: they show where the platform is heading, not what it does today.
Architecture: what sets it apart
If every application talks directly to every model, the integrations grow with the product of the two numbers. With a shared knowledge layer they grow with their sum.
Applications and channels
Models and providers
3 applications and 3 models: 9 direct integrations, 6 with the shared layer.
One more model costs 3 direct integrations, or a single one with the shared layer; one more application costs 3, or one.
Web, Telegram, exports. A new channel reads the same knowledge layer; the architecture is also designed to support a Windows application.
A catalog of models and providers with a policy for each customer. Models are swapped through configuration, and the policy decides which ones are allowed.
Jobs declare what they need, workers declare what they can do. They can run on the central server or on remote nodes.
Documents, versions, chunks with provenance and indexes. It is the one part that stays put while the others change.
Preview
Every knowledge base will be able to generate its own study material and the quizzes to test it, with every statement linked to the original documents.
Design complete, implementation in progress. Features, timing and details may change before release.
An agent reads the knowledge base and proposes the chapter outline. The training lead confirms it.
Text with tables and diagrams, where every statement cites its source.
Deterministic checks and a second agent make sure citations and content match.
Multiple-choice questions with four options and one correct answer, at three levels: recall, understanding, application.
At the end of the quiz every mistake points to the original document and to the study unit to review.
Preview · example content
Unit 3 · Level: understanding · Question 1 of 5
Pick an answer to see what reviewing a mistake will look like.
Example content, not the final interface.
Knowledge in different contexts
What sets a contract apart from course notes or tax guidance is not in the code. It is in a domain profile: a versioned JSON file that describes structure, references and rules. A new domain needs no new code.
Domain profile
domain-packs/generico/profile.json
{
"code": "generico",
"language": "it",
"chunking": {
"strategy": "recursive",
"targetTokens": 380,
"maxTokens": 480,
"overlapTokens": 40
}
}A simplified excerpt of a real profile. The same engine reads all four: only the file changes.
Trust, control, traceability
Accuracy is proven through provenance. Select a citation to see what ELYVRA keeps with every answer.
Example · fictitious content
Answer generated by artificial intelligence
Question · How long does the product warranty last?
The warranty lasts twenty-four months from delivery. For wear parts the period is reduced to twelve months.
“The warranty lasts twenty-four months from the date of delivery.”
Every answer keeps the document versions, profile and parameters it used, so you can reconstruct how it was produced.
Content is never overwritten. A superseded document leaves search on the day the superseding document takes effect.
Filters in the application and row-level security in the database. Authorization comes before retrieval.
Each customer chooses which providers to allow. If no model complies, there is no fallback to a non-compliant one.
Optional pseudonymization: the mapping between placeholders and data stays in memory only for the duration of the request.
Roles with permissions, two-factor authentication for administrators, a log of administrative operations.
Customer data is stored in the European Union. An AI provider outside the EU receives text only if the customer's policy allows it, with an explicit, recorded confirmation.
Request a demo
ELYVRA is in pre-launch, available on request for teams who want to try it. Describe your use case: the request is sent from your own email app; this page sends nothing.
Or email us directly at
Elevating information to knowledge
Hello, I would like to find out whether ELYVRA fits this use case. Sources: Documents Destinations: Chat with citations Knowledge: to be described Thank you.