Build AI-powered Q&A replies with a knowledge base
Upload documents, FAQs, or product info, and the AI Conversation node uses them to ground every reply.
What this is for
The AI Conversation node responds to inbound messages using GPT. By default, it uses just the system prompt you configure. With a knowledge base attached, the AI can reference your actual product info, return policy, FAQs, or any document you upload, and ground answers in real content rather than guessing.
Prerequisites
- An OpenAI API key configured under Settings → AI Settings
- A connected WhatsApp number on Walytic
- A flow with an AI Conversation node (see Build your first flow)
Step 1: Create a knowledge base
Sidebar → Knowledge Bases → New Knowledge Base.
Upload one or more files: PDFs, plain text, or markdown. Large documents are chunked automatically. Walytic builds an embedding index in the background. For most knowledge bases, indexing finishes within a minute or two.
Tip: smaller, focused knowledge bases work better than one giant kitchen-sink document. Keep separate KBs for "product FAQ", "shipping policy", and "return policy" rather than one combined file.
Step 2: Attach the knowledge base to a flow node
Open your flow. Click the AI Conversation node. In the configuration panel:
- Prompt: keep it concise. Tell the AI its role, tone, and any escalation rules.
- Knowledge bases: pick one or more from the dropdown. The AI gets the relevant chunks of those documents injected into every conversation.
Save the flow.
Step 3: How it answers
When a customer message arrives at the AI Conversation node:
1. The customer's message is embedded.
2. Walytic finds the most relevant chunks from the attached knowledge bases.
3. Those chunks are injected into the system prompt: "Knowledge base: …".
4. GPT generates a reply that references the chunks.
The reply is sent back to the customer via your WhatsApp transport. The conversation can continue (the node maintains context across turns) until the customer types your exit keyword (e.g., "stop") or hits the max turns limit.
Tips for better answers
- Be specific in the prompt. "You are a customer service assistant for Acme Stores" is better than "You are helpful." Mention your tone (formal, friendly), language preferences, and what to do if the question is out of scope.
- Set a max turns. Without it, conversations can spiral. 5-10 turns is typical.
- Update the KB regularly. Re-upload when policies change. Old documents continue to be indexed until you remove them.
- Test with edge cases. Send the kind of weird, off-topic, or rude messages that real customers send. The AI behaves differently on those than on your "happy path" tests.
- Monitor cost. GPT-4o-mini is cheap, but a chatty conversation across thousands of users adds up. Use the AI usage dashboard under Settings to track.
Pricing
The AI Conversation node uses the OpenAI key you configure. You're billed by OpenAI directly for tokens consumed. Walytic does not mark up. Typical cost: a few cents per multi-turn conversation with GPT-4o-mini.
Common questions
Can I use Anthropic, Gemini, or another provider?
Currently OpenAI only. Anthropic support is planned.
Does the knowledge base see customer messages?
The KB is read-only for retrieval. The AI sees customer messages and KB chunks; the KB itself is not modified.
Can I make the AI hand off to a human?
Yes. Set the exit keyword (e.g., "human") and configure a follow-up flow that pings your team's Unibox when triggered.
Need more help?
Browse the rest of the help center or contact support directly.