Skip to main content

Knowledge bases and RAG

A Wirevox knowledge base is a workspace collection of indexed source material. Retrieval-augmented generation, or RAG, searches that material for content related to a person's question and gives the relevant results to the conversational model before it answers.

Use a knowledge base for approved information that is too large, detailed, or frequently maintained to place in a system prompt: policies, service descriptions, product details, troubleshooting procedures, FAQs, and internal reference documents.

RAG does not train the selected agent model or permanently teach it the source material. It retrieves relevant source chunks at conversation time. The model still needs clear instructions for grounding, uncertainty, and escalation.

In this topic​

Choose the correct context layer​

NeedUse
A compact stable value such as business hoursPrompt variable
A behavioral rule or boundarySystem prompt
Detailed approved reference materialKnowledge base
Current external state or a write operationFunction or integration
Fresh public-web researchA web research tool, then an explicit approved save if it should become knowledge

Do not duplicate the same fact across all four layers. Choose one authoritative location and define what happens when it has no answer.

Workspace and agent boundaries​

Knowledge bases belong to a workspace and can be reused. An agent-level configuration selects one workspace knowledge base. A reusable Knowledge function can search a whole base or an exact subset of its sources and can be attached to voice or chat agents.

Knowledge changes and agent publication are separate. Adding or removing source material changes the shared indexed collection. Attaching a base or Knowledge function is an agent draft change that must be tested and published for live agents.

What RAG does not guarantee​

Retrieval improves access to approved information, but it does not guarantee that:

  • the source was accurate or current;
  • a website or file was extracted completely;
  • the most relevant chunk was returned;
  • conflicting sources were resolved correctly; or
  • the model will follow weak or contradictory grounding instructions.

Source governance, realistic evaluation, and a safe “not found” path remain necessary.

Read Understand Wirevox RAG, then create a small collection with representative questions before importing a large site or document library.