Applications and AI · 04

RAG for business: ask your documents and check the source.

An employee asks a question, the knowledge base finds the relevant passages of documents, and a language model writes the answer with a source reference. That is RAG. Everyone uses only approved materials they have permission to access.

01

Spend less time searching scattered materials

For teams that regularly ask about procedures, products or decisions recorded in documents. Below are three illustrative use cases.

Employee procedures

A question about submitting an expense leads to the relevant passage in the current procedure. The employee can open the source and check conditions and exceptions before taking action.

Technical documentation

A person handling a support request finds the instructions for the correct product variant. The answer identifies the material it relies on and should not mix information from incompatible versions.

Materials for a proposal

A salesperson asks about approved specifications, service scope or terms recorded in accessible documents. They verify the cited passages before using the answer in a client conversation.

02

Knowledge, sources and access rules

Preparing documents

We agree on sources, formats, versions and content owners. We design document ingestion, splitting into passages and index updates. Whether documents can remain in their current location depends on the integration available.

Answers with references

We connect retrieval with the model and a source view. The user can see the document and the passage used in the answer. We define behaviour when information is missing or contradictory.

Permissions and operation

We design access controls for retrieval and opening sources. We agree on logging, retention and responsibility for keeping knowledge current, removing documents and changing permissions.

03

Start with a selected body of knowledge

We need a representative sample of documents, a list of typical questions and the location of the correct answers. Include questions the collection cannot answer, outdated materials and documents available only to specific roles.

We agree on the knowledge owner, authoritative source, frequency of changes and how access is inherited. Scan quality, tables, duplicates and missing version labels may require additional preparation. Confidential files should only be shared through an agreed channel.

04

A pilot using your team's questions

  1. 01

    Select the collection and roles

    We define documents, users, expected answers and cases where access must be denied or the available information is insufficient.

  2. 02

    Prepare retrieval

    We build the index and its update mechanism. We check whether questions lead to the correct passages and document versions.

  3. 03

    Test the answers

    We assess consistency with the source, the usefulness of references and compliance with permissions. Users report incorrect, incomplete and misleading results.

  4. 04

    Check changes to the knowledge base

    We update and remove documents, change access and repeat the tests. We agree on deployment conditions and responsibility for maintaining the collection.

05

The source must actually support the answer

We check that the answer matches the cited passage, not just that a link is present. We test questions without a source, conflicting versions and outdated materials. We also check that document content does not override the assistant's agreed operating rules.

Tests across different roles cover retrieval, answers and opening sources, including after access has been revoked. We separately check document updates and removal from the index. Quality criteria and refresh times are agreed for the specific collection; we do not promise an answer to every question.

06

What affects the cost of a knowledge base

The number of sources, document quality, update frequency and complexity of permissions all matter. Model selection, infrastructure, interface and testing also contribute. The estimate separates collection preparation and implementation from ongoing maintenance. We define scope and timing after reviewing the materials.

FAQ

Frequently asked questions

Does RAG mean training the model on our files?

No. In this approach, relevant passages are retrieved and supplied to the model as context for a question. Model training is a separate subject and is not included in this service by default.

What happens if the documents do not contain an answer?

We design a response for insufficient information and test it on such questions. The model can make mistakes, so this behaviour is part of the quality assessment, not just a stated intention.

Will every employee see the entire knowledge base?

We design access around agreed roles and source permissions. We check it during retrieval, answer generation and document opening, including after permissions have changed or been revoked.

Does the website demo search our documents?

No. The assistant on Entrada's website demonstrates a conversation with a model and has no access to your files. We build a company RAG knowledge base separately, using an agreed collection and environment.

→

Related services

Which documents does your team ask about most often?

Tell us where the knowledge is stored, who uses it and which questions keep coming up. We will agree on a collection for the pilot and how to assess the answers.

Prefer email? biuro@entrada.pl