How Edah AI Stays Grounded: Knowledge Bases and Retrieval in Voice

Why voice AI hallucinates

A generic language model predicts plausible text. For a chatbot, a plausible-sounding wrong answer is annoying. For a voice agent, it is a customer who hangs up and a brand that loses trust. Voice AI cannot afford the same hallucination tolerance as a chatbot, because the customer hears the answer in real time and reacts to it on the call.

How grounding works

The agent does not reason in a vacuum. Before generating any response that touches your business data, it retrieves relevant context from your knowledge base. Product details, pricing, hours, FAQ, policies, the things only you can confirm. The retrieved context shapes the response. The model is not allowed to invent.

Chunking and embedding

Your knowledge sources, whether they are PDFs, Excel sheets, Markdown documents, or web pages, get chunked into semantic units and embedded into a vector store. When a question comes in, the agent retrieves the most relevant chunks and feeds them into the reasoning step.

Reranking and confidence

Not every retrieved chunk is equally relevant. A reranking step orders the candidates by actual relevance to the specific question. If no chunk crosses the confidence threshold, the agent refuses to answer from memory and either says it does not know or hands off to a human. Silence beats a wrong answer.

Source attribution

Every grounded answer can be traced back to the chunk it came from. In the audit log, the chunk identifier, the source document, and the retrieval score are all recorded. The reviewer can see exactly what the agent based the answer on.

What this means in practice

The customer who calls and asks "is the deluxe suite still available with breakfast included" gets an answer pulled from your live availability source and your current rate sheet, not from a model guess. The accuracy compounds across thousands of calls. The brand trust holds.