AI assistant on company knowledge
How do we build an AI assistant that answers from our company's own documents?
The short answer
TEVEL (תבל) builds knowledge assistants and AI agents that answer from the company's own documents and systems using RAG — retrieving relevant information from approved sources first, then composing an answer that cites the source. The difference between an impressive demo and a tool people can trust lies in permissions, source selection, logging and a clear definition of what the assistant does not do.
How it works
RAG (retrieval-augmented generation) connects a language model to the organization's knowledge base: documents are processed and indexed for semantic search, and for every question the system retrieves the relevant passages and answers based on them. The answer relies on company knowledge rather than only the model's general knowledge, and can show which document it came from.
What makes it trustworthy
Principles defined before any code is written:
- Approved sources only — what goes into the knowledge base and who keeps it up to date
- Permissions — each user gets answers only from information they're allowed to see
- Source references on every answer, so it can be checked
- A transparent response when there isn't enough information, instead of guessing
- Logging of questions, answers and sources for oversight and improvement
- Actions (if any) defined up front and requiring approval where needed
Possible uses
An internal copilot for staff (procedures, policies, product knowledge), an assistant that suggests answers to service agents from approved knowledge, search across a document archive, or a knowledge layer inside a customer-service operation. Always inside the systems people already work in — not as yet another separate tool.
More questions
Is our data used to train the model?+
System architecture, model providers and data-use terms are chosen to fit the organization's requirements. With RAG, information is retrieved at question time and isn't needed for model training.
What happens when the assistant doesn't know?+
It's set up to say so explicitly and refer to a person, rather than guess.
Can it connect to our CRM or ERP too?+
In many cases yes, subject to APIs and permissions, so the assistant can retrieve live data and not only documents.