n8n support bot with grounding cuts fabricated answers from 22 to 0

On November 6, 2026, a developer posted on DEV Community an account of running a before-and-after comparison for a customer support FAQ workflow in n8n. The test setup was a self-hosted n8n 1.123.84 instance backed by Qwen 1.8B (Q4_K_M) loaded through Ollama on CPU, paired with a knowledge base of 6 entries. The author asked 20 questions: 10 that the knowledge base did cover and 10 that it did not, then asked every question three times, against both a "raw" version and a "grounded" version, for 120 runs in total. The raw configuration produced 22 answers that did not appear anywhere in the knowledge base, while the grounded configuration produced zero fabricated answers, with average latency around 1.4 seconds. The developer was explicit about the limits of the evidence, noting that the test set was small and the conclusions applied only to this particular support FAQ scenario rather than to broader question answering. The post opens with a short framing remark: a bot that always answers "I don't know" is frustrating, but a bot that confidently insists "No, we don't have a phone number" when nothing in the documentation ever said that is a liability, which is what motivated the author to test their own workflow. As of the available reporting, no follow-up articles or independent verification have appeared, so the episode remains a single developer experiment rather than a documented change in n8n or Ollama behavior.