How does the AI knowledge base work?

Upload PDFs, docs, URLs and FAQs. VoiceForge chunks, embeds and retrieves with RAG at call time — the agent answers from your sources, cites them, and falls back to transfer when confidence is low.

Does it hallucinate answers not in the docs?

No. Retrieval is grounded: answers come from retrieved chunks with a confidence threshold. Below threshold the agent says it does not know and offers a warm transfer with context instead of guessing.

Can each client have a separate knowledge base?

Yes. Each workspace has an isolated knowledge base with its own documents and embeddings. No cross-client leakage — essential for agencies reselling to multiple businesses.

How is this better than pasting FAQs into the prompt?

Prompts truncate and cannot cite sources. RAG retrieves only relevant chunks per turn, stays within context limits, updates without republishing the agent, and logs which chunks grounded each answer.

Grounded answers, not guesses

Knowledge base that answers from your docs — with citations

Upload the docs, URLs and FAQs you already have. At call time the agent retrieves only relevant chunks, answers with citations, and transfers when confidence is low.

Start building See how it works

RAG at call time

Documents are chunked and embedded on upload. Each turn retrieves top-k chunks by relevance — the agent answers only from what was retrieved, with source tracking per answer for QA.

Citations, not hallucinations

Every grounded answer can cite its source chunks. Below the confidence threshold the agent discloses it does not know and offers warm transfer with the caller context attached.

Update without republishing

Add or replace documents and the next call retrieves from the new version. No prompt rewrite, no agent republish — the knowledge base is versioned independently from the Agent Spec.

Per-workspace isolation

Each client workspace has its own knowledge base and embeddings. No leakage across clients, no shared index — the isolation agencies need to resell confidently.

Pairs with booking and transfer

FAQ resolved via RAG, booking via calendar tools, escalation via warm transfer — one declarative agent handles all three without custom wiring per intent.

Frequently asked questions

How does the AI knowledge base work?: Upload PDFs, docs, URLs and FAQs. VoiceForge chunks, embeds and retrieves with RAG at call time — the agent answers from your sources, cites them, and falls back to transfer when confidence is low.
Does it hallucinate answers not in the docs?: No. Retrieval is grounded: answers come from retrieved chunks with a confidence threshold. Below threshold the agent says it does not know and offers a warm transfer with context instead of guessing.
Can each client have a separate knowledge base?: Yes. Each workspace has an isolated knowledge base with its own documents and embeddings. No cross-client leakage — essential for agencies reselling to multiple businesses.
How is this better than pasting FAQs into the prompt?: Prompts truncate and cannot cite sources. RAG retrieves only relevant chunks per turn, stays within context limits, updates without republishing the agent, and logs which chunks grounded each answer.