Prepare the knowledge base before configuring chat. This workflow searches existing Qdrant documents; it does not crawl or index your website.
Import the JSON and leave it inactive. Assign OpenAI credentials to classification, smalltalk, embeddings and the answer-model nodes. Assign Cohere credentials to Cohere Rerank by Relevance.
In Search Qdrant Knowledge Base, connect your instance and replace website_knowledge with your populated collection. Confirm stored vectors are compatible with the query embedding model.
In When Chat Webhook Received, keep POST and path website-ai-chatbot, and configure your own Header Auth credential.
Use the test URL for manual validation. The website backend should hold the private header and forward browser messages; keep it out of public JavaScript. Configure input limits, rate limiting and provider-error handling at the application or gateway.
Send an unrelated request: reject should return a controlled text reply without retrieval. Send a greeting: smalltalk should produce valid JSON with type: text and a non-empty text field.
Ask about an indexed page. Search should return relevant titles, URLs and summaries, with no more than five results. Test a question without usable matches and confirm an empty results array with an appropriate follow-up rather than invented sources.
Test missing/invalid authentication and unavailable providers. Inspect actual HTTP statuses and response shapes; the template does not standardize every provider failure.
Verify the frontend handles text, links, empty results and errors using 02 — Webhook API & Responses. Activate the workflow, switch the backend to its production webhook URL and repeat a controlled request.
Review latency and usage before scheduling wider use. The supplied retrieval starts with 50 Qdrant candidates and 20 reranked candidates; tune it after the basic route works.