I've looked a bit into this. I've also converted the function into a tool (although I was pretty confident that it was not the issue). For a tool response such the one below:

{
  "name": "get_restaurants",
  "role": "tool_call_result",
  "time": 1722870773304,
  "result": "Zia Lucia; Brasserie Blanc; Duck & Waffle; New Street Grill; Tinseltown; Emilia's Crafted Pasta; Cabotte; Barcelona Tapas Bar & Restaurant",
  "toolCallId": "call_kkxHmJO2KO9c2qxnwvgbzWOL",
  "secondsFromStart": 68.439

The LLM generates this message:

{
  "role": "bot",
  "time": 1722870773417,
  "source": "",
  "endTime": 1722870805967,
  "message": "Hold on a sec. Here are some excellent dining options in the Aldgate East area. 1, known for its delicious pizzas with a variety of dough options. 2, offers classic French cuisine in a relaxed setting. 3, a unique dining experience with British and European dishes located high above the city with stunning views. 4, a great spot for steak and seafood lovers. 5, an American style diner with a wide range of comfort food. 6. Specializes in fresh handmade pasta dishes.",
  "secondsFromStart": 68.29
}

Where the names of the restaurants are not mentioned. Additionally, it seems to pick additional information from somewhere else. Those information were provided by a previous version of our tool, so I'm suspecting some sort of caching is going on.