From corporate documents to dependable AI answers
AYITA introduces Knowledge Base and ARAG, two connected capabilities designed to help enterprise AI work with an organization’s own information. The Knowledge Base decides which documents AI is allowed to use. ARAG - AYITA’s engine for finding and researching evidence - searches those documents, gathers the relevant passages, and builds answers that cite them. Together, they address a challenge that goes far beyond searching files: how can a company make its knowledge useful to AI without losing control of who may use it, which version is authoritative, and what evidence supports an answer?
An employee asks, “What do our policies require before we share customer data with a partner?” Another wants to compare two regulatory obligations. A project team needs to find a decision buried across specifications, meeting notes, and internal guidance.
The answers may already exist somewhere in the company. Finding and reconciling them can take hours. An AI assistant could make this work dramatically easier: ask in ordinary language, receive a clear explanation, follow the sources.
But corporate knowledge is not one large pile of text. It is spread across formats, departments, languages, and versions. Some documents are approved; others are drafts. Some are broadly available; others are restricted. Some contradict each other because one has superseded another.
So the real question is larger than “Can AI find a relevant paragraph?” Can an organization trust the way AI works with its knowledge?
That question guided the design of these capabilities: the Knowledge Base governs the documents; ARAG retrieves and researches their content; and the resulting answers remain connected to their sources and access rules.
Turning documents into governed knowledge
Traditional search locates files. An AI assistant goes further: it interprets passages, connects information across documents, and produces an answer. To make that useful inside a company, the content must be discoverable, current, authorized, and traceable all the way to the answer.
The AYITA Knowledge Base prepares documents for both people and AI. It imports common enterprise formats, maintains revisions, and tracks extraction quality. But ingestion alone does not make a document available to AI. A Data Steward reviews and approves it, it is published, and then it is separately released for RAG use. These are deliberate steps in controlling which corporate knowledge the assistant may draw on.
From upload to citation
How one document becomes evidence an AI answer can cite, and how every step stays under the workspace's rules.
- Text and layout extracted
- Recorded as revision 3
- Extraction quality checked
- Imported
- Approved by Data Steward
- Published to the workspace
- Released for AI
Can we share customer data with a partner for joint analytics?
- Checking access
- Searching the workspace
- Researching
- Checking each claim
Yes, with conditions. A signed data processing agreement must be in place before any transfer 1, and the partner may use the data only for the agreed purpose 2.
Completed · sufficient evidence4.2 Disclosure to third parties. Customer personal data may be disclosed to a partner only after a data processing agreement has been signed by both parties. The agreement must name the purpose of processing and the retention period.
Upload
Illustration. The document, question, and answer are examples; real workspaces, access rules, and timings depend on the deployment.
ARAG works with that prepared knowledge. It combines semantic search, keyword retrieval, and exact identifier matching, then uses the relevant evidence to support answers or more extensive research.
This matters in practice. A compliance analyst might ask in English about a provision in a document written in another language. A legal team might search for a precise article number. An engineer might describe a problem without knowing the internal terminology used in its specification. Different questions demand different ways of finding the evidence.
In development testing, we evaluated ARAG against 211 questions in two languages, including cross-language queries, covering 24 regulatory and standards documents. The correct document was retrieved for 98.8% of questions, and the relevant passage appeared among the top 20 results for 93%. These are measured results for that defined test set, not claims about every possible corporate corpus.
Finding evidence, however, is only the first step.
An AI assistant should not outrank the company’s rules
Imagine that an employee is cleared to read a policy summary but not its confidential supporting analysis. For that employee, the higher-classified document should effectively not exist in the AI’s available knowledge - not merely appear as a result that is hidden at the last moment.
In AYITA, classification and the user’s clearance constrain what can be retrieved, while the Knowledge Base remains responsible for document permissions and lifecycle. ARAG’s search index is a derived working copy, not a new source of authority. Verified identity, tenant, and workspace establish the boundary; authorization is checked before evidence is released. Search parameters cannot widen access, and evidence is withheld if permission cannot be verified.
Now suppose the document owner withdraws a policy. The withdrawal must affect what the AI can use, not merely what appears in a file browser. AYITA makes withdrawn material unavailable to retrieval before cleaning up its indexed copies and cached evidence. The design target is removal from searchable results within five seconds of ARAG receiving the change event. The precise timing will depend on the deployment.
AI does not acquire a permanent right to use information simply because it indexed it yesterday.
Trust is something a user can examine
Even an authorized, up-to-date document does not automatically produce a correct interpretation. That is why ARAG does more than retrieve passages and ask a model to write a fluent response.
For research questions, it can plan several retrieval steps, gather evidence, develop an answer, and check claims against supporting passages. Citations lead to specific document revisions and locations, so a reviewer can inspect the basis for an answer instead of being asked to accept it on authority.
Just as importantly, the system distinguishes a completed, evidence-supported result from one where evidence is insufficient, a time or token budget was exhausted, or the user canceled the investigation. Those outcomes are visible in AYITA’s Insight panel.
Consider the difference between two responses to a difficult question:
“Here is the answer,” followed by a convincing paragraph.
Or: “Here is what the available documents support, here are the passages, and here is what we could not establish.”
The second may sound less magical. In an organization where people are accountable for decisions, it is far more useful.
Designed for the enterprise, not just the demo
The Knowledge Base and ARAG are part of AYITA, Realm’s governed AI runtime. The components are designed to operate within a customer’s environment, using self-hosted models through AYITA’s inference gateway, without sending documents to an external AI service for this workflow.
The architecture also anticipates ordinary operational realities: indexes need rebuilding, documents are republished, workers fail, searches run while content changes, and users stop tasks before they finish. Preserving the rules under those conditions is part of making the assistant dependable.
What a knowledge or compliance team should test
The future of enterprise AI is not simply a better chat interface placed on top of corporate files. It is AI becoming a reliable participant in how organizations find, interpret, and use what they know - with the same ownership, permissions, versioning, and accountability companies already apply to their information systems.
For a compliance, legal, or knowledge-management team, the test is not whether an assistant can produce a fluent answer. It is whether the organization can control which knowledge the AI uses, trace every answer to an authorized source, and see where the evidence runs out. A controlled pilot on one body of knowledge should measure both the time saved and the quality of that evidence:
- time to find and reconcile an answer, compared with today’s manual search;
- share of answers whose citations lead to the right document revision and passage;
- answers that drew on material beyond the user’s clearance - the target is zero;
- time for a withdrawn document to disappear from answers;
- questions where the system stated what it could not establish - and whether that was the right call.
The question is no longer only whether AI can answer. It is whether the organization can understand, verify, and trust how that answer was produced.
Discuss one body of knowledge with Realm: pick a policy set or a regulatory library, identify who may use it, and define the evidence your reviewers need behind every answer.
Retrieval metrics and operational timings quoted here come from AYITA development testing. They are specific to the tested corpus and environment and do not represent universal performance or compliance guarantees.