A voice agent never has its own knowledge base. It borrows everything from a Vex AI chatbot you link to it. That means the same articles, FAQ snippets, and custom tools that power your chat experience also power your phone calls — and any improvement you make to one channel improves the other automatically.
One brain, two mouths
When you link a Vex chatbot to a voice agent on the Connections tab, the voice agent inherits three things from that chatbot:
- Knowledge sources — KB articles, FAQ snippets, scraped public URLs, uploaded documents, and product catalog entries.
- Workflow tools — custom tools you defined on the chatbot, such as fetching order details, creating tickets, or calling an external API.
- Context tools — built-in tools that read the caller's phone number, the current call duration, your workspace business hours, and the agent's own configuration.
The voice agent's system prompt is its own — it does not inherit the chatbot's prompt. Only knowledge and tools cross over.
How to link a Vex chatbot
- Open your voice agent from Dashboard → Automation → Voice.
- Switch to the Connections tab.
- Under Vex chatbot, pick the chatbot whose knowledge you want the voice agent to use.
- Save.
The link is live immediately. The next call the agent takes will draw from the linked chatbot's knowledge.
How the agent decides what to use
For each caller turn, Gemini Live decides whether to answer directly or call a tool. In practice:
| Caller asks | What the agent does |
|---|---|
| A general factual question ("what are your shipping zones?") | Knowledge base lookup |
| A personal or account question ("where is my order?") | Workflow tool that looks up the order number |
| A procedural question ("are you open right now?") | Context tool that checks business hours, or knowledge base if you have an article about hours |
| Small talk or something covered in the system prompt | Direct answer, no tool call |
The agent often says "let me check that for you" when it needs a beat to call a tool. That latency is intentional — it tells the caller the agent is doing real work rather than guessing.
What makes good voice knowledge
The same articles work for chat and voice, but a few practices help on the phone specifically:
- One canonical answer per topic. If three articles overlap on the same question, the agent hesitates while picking the best one. Merge them.
- Spell out abbreviations. The agent will read what you wrote — "USPS" sounds like four letters, but "U.S. Postal Service" sounds natural.
- Keep answers spoken-length. A 500-word article is great for the help center but reads as a monologue on the phone. Lead with a one-sentence summary; the agent will use that on the call.
- Avoid bullet-heavy lists. The agent will literally say "first" or "second" if the structure is too rigid. Smooth prose flows better when spoken aloud.
If callers complain that the agent is too slow on basic questions, the knowledge base is usually too fragmented — many short overlapping articles competing for the same query. Consolidate into one canonical article per topic and the hesitation disappears.
Updating knowledge for both channels
Because there is one shared brain, updates apply everywhere at once. Add a new article to your knowledge base and:
- The chatbot picks it up on the next message.
- The voice agent picks it up on the next call.
You never need to maintain a "voice copy" of an article. See Set up Vex AI for the first time for how to add knowledge sources to a chatbot.
What to read next
Now that the agent has knowledge to work with, configure how it behaves after hours: Business hours for a phone number.