RAG-Powered AI Voice Customer Support Agent (Supabase + Gemini + ElevenLabs)

Execution video: Youtube Link

I built an AI voice-triggered RAG assistant where ElevenLabs’ conversational model acts as the front end and n8n handles the brain....here’s the real breakdown of what’s happening in that workflow:

1. Webhook (/inf)

   • Gets hit by ElevenLabs once the user finishes talking.
   • Payload includes user_question.

2. Embed User Message (Together API - BAAI/bge-large-en-v1.5)

   • Turns the spoken question into a dense vector embedding.
   • This embedding is the query representation for semantic search.

3. Search Embeddings (Supabase RPC)

   • Calls matchembeddings1 to find the top 5 most relevant context chunks from your stored knowledge base.

4. Aggregate

   • Merges all retrieved chunk values into one block of text so the LLM gets full context at once.

5. Basic LLM Chain (LangChain node)

   • Prompt forces the model to only answer from the retrieved context and to sound human-like without saying “based on the context”....
   • Uses Google Vertex Gemini 2.5 Flash as the actual model.

6. Respond to Webhook

   • Sends the generated answer back instantly to the webhook call, so ElevenLabs can speak it back.

You essentially have:
Voice → Text → Embedding → Vector Search → Context Injection → LLM → Response → Voice