Troubleshooting

Common issues and fixes

Use this page when a deployment, provider integration, MCP workflow, redaction chain, or portal build does not behave as expected.

Deployment stays in Provisioning

Likely cause: Provider-backed deployments should become ready once the inference target is bound. If they do not, the binding or API rollout may be stale.

Fix: Open the deployment detail, confirm it has an inference target, then restart or redeploy the API if provider readiness backfill has not run.

Model node has no content to send

Likely cause: The model node prompt_field points at a payload field that does not exist.

Fix: Use a field that exists after upstream nodes. If redact_text edits raw_tickets.text in place, use prompt_field: raw_tickets.text. If you create a top-level redacted_text field, use prompt_field: redacted_text.

Claude or another provider does not receive MCP data

Likely cause: The mcp_input result_field or model prompt_field is mismatched, or the MCP tool returned only structured data without text.

Fix: Check that mcp_input writes a stable result_field and that the model reads <result_field>.text. Quantlix normalizes MCP output into text, raw, items_count, and has_text.

Customer names or phone numbers appear in model output

Likely cause: The redaction patterns did not match the content, or the model read raw context before redaction.

Fix: Run redact_text before model, add extra_patterns for international phone formats, and inspect What reached Claude in the audit log.

Portal dev server fails with EADDRINUSE on port 3001

Likely cause: Another Next.js dev server is already listening on port 3001.

Fix: Stop the old process or run pnpm dev -- -p 3002.

What to collect before asking for help

  • The workflow ID, deployment ID, and execution ID.
  • The node that failed and the exact error message.
  • The relevant node config, especially result_field, fields, and prompt_field.
  • The audit log section showing what reached the provider model.