The embeddings generated from a single sentence will be different than the embeddings generated for an entire chunk/node. The node has more context, which biases the embeddings, compared to a single sentence
Hence, typing exact sentences is not a good way to test vector similarity. It's more about the similarities in overall semantics.
You can augment vector search by also using a keyword index. There is an example here that sets up a custom retriever to do just that
https://github.com/jerryjliu/llama_index/blob/main/docs/examples/query_engine/CustomRetrievers.ipynb