Grounded RAG Support Ticket Bot with Claude, Pinecone & n8n

Grounded RAG support replies—instantly—powered by Claude, Pinecone, and n8n

This n8n workflow receives customer support questions via a webhook, retrieves the most relevant documentation from Pinecone using Voyage AI embeddings, and uses Anthropic Claude to generate a grounded, cited answer. It can auto-reply when confidence is high or seamlessly escalate to Zendesk and Slack when it’s not.

What this workflow does

• Ingests support requests through an HTTP webhook, keeping the connection open for a synchronous response.
• Normalizes the payload (query text, customer/ticket identifiers, and product line) and generates an embedding using Voyage AI.
• Retrieves relevant documentation from Pinecone within the matching namespace, then deduplicates/ranks results and builds a numbered context block with retrieval confidence scores.
• Generates a grounded JSON response with Anthropic Claude, including citations to the retrieved sources.
• Validates citations and decision logic, checking confidence and escalation keywords before acting.
• Auto-resolves or escalates:
  • When qualified, replies to the webhook with the grounded answer, citing sources, and tags the ticket as auto-resolved in Zendesk.
  • When retrieval is weak or escalation is required, opens/updates a Zendesk ticket, notifies the team in Slack, and returns an escalation notice to the webhook.
• Logs every interaction—queries, retrieval metrics, decision path, and sources—into Google Sheets for auditing and continuous improvement.

Use cases

• Automating level-1 SaaS support with retrieval-augmented generation (RAG) and cited answers.
• Reducing response time by auto-answering high-confidence questions while routing uncertain cases to human support.
• Maintaining compliance/traceability with auditable logs and source citations.

Technical details

• n8n nodes/tools: webhook, if, code, merge, slack, zendesk
• RAG stack: Voyage AI embeddings → Pinecone retrieval → Anthropic Claude grounded generation
• Auditing: interaction logs to Google Sheets