Spinning up a FastAPI endpoint to intercept Cognigy.AI utterances, run them through sentence-transformers/all-MiniLM-L6-v2, and push ranked snippets back via POST /webhook/v1/execute before the NLU classifier kicks in. Payload shape looks fine locally, but the ctx.data object keeps getting flattened when it crosses into the Cognigy runtime, so ctx.data.knowledgeChunks ends up empty even though the Milvus search() call returns valid cosine scores. We’ve tried swapping the JSON serialization to json.dumps(response, default=str) and verified the Content-Type: application/json header, but the webhook timeout hits 408 before the embedding pipeline finishes.