Text flow broken in two step generation with function calls + text

Gemini API
, ,
1

Hi everyone,

I am building a conversational agent using the Node SDK (@google/generative-ai). I am running into a persistent issue where the model kind of repeats itself in the second step after receiving a functionResponse, instead of simply continuing the previous generated text.

In practice, first step text ends with tool response introduction (“Here is a selection of …:”), but the second step text also starts with these data introduction ('Here are some…).

The Workflow:

  • Step 1: User asks a question (e.g., “How to choose a pair of shoes?”).
  • Step 1 Output: The model streams text advice AND a functionCall (e.g., suggestProducts) in the same generation.
  • Step 2: Function is executed client-side.
  • Step 2 Input: I construct the full history (User → Model → FunctionResponse) and run a new generation to present data.

Here is a sample context input , which has been generated with gemini api (generateContent with function calls).

[
  {
    "role": "user",
    "parts": [
      {
        "text": "Can you explain how to choose running shoes and then find me some highly cushioned ones?"
      }
    ]
  },
  {
    "role": "model",
    "parts": [
      {
        "text": "Choosing the right pair of running shoes is essential for comfort and preventing injury. Here are key factors to consider:\n\n1.  **Terrain**: Are you running on the road, on trails, or on a treadmill?\n    *   **Road shoes**: Designed for pavement with good cushioning to absorb impact.\n    *   **Trail shoes**: Offer better grip and protection for uneven surfaces.\n2.  **Pronation**: This is how your foot rolls when you run.\n    *   **Neutral**: The foot rolls slightly inward. Most runners fall here.\n    *   **Overpronation**: The foot rolls inward excessively. You might need stability shoes.\n3.  **Cushioning**: Ranging from barefoot feel to maximum cushion. Maximum cushion is great for long distances and joint protection.\n4.  **Size**: Feet swell during runs, so ideally, leave a half thumb's width of space at the front of the shoe.\n\nHere is a selection of running shoes with maximum cushioning for your comfort:\n\n"
      },
      {
        "functionCall": {
          "name": "suggestProducts",
          "args": {
            "toAvoid": [
              "minimalist",
              "racing flats",
              "hard sole",
              "barefoot"
            ],
            "universe": "sport",
            "minPrice": 80,
            "exclude": [],
            "descriptions": [
              "highly cushioned road running shoes",
              "max cushion sneakers for long distance",
              "comfortable running shoes shock absorption"
            ],
            "maxPrice": 250,
            "categories": [
              "running shoes",
              "sport shoes"
            ],
            "targetPrice": 150
          }
        },
        "thoughtSignature": "..."
      }
    ]
  },
  {
    "role": "user",
    "parts": [
      {
        "functionResponse": {
          "name": "suggestProducts",
          "response": {
            "products": [
              {
                "id": "shoes-run-cloud-101",
                "name": "CloudStratus Max",
                "description": "Double layers of CloudTec® for maximum cushioning on the road. Perfect for medium to long runs.",
                "price": "169.95€",
                "url": "https://example.com/products/cloudstratus-max"
              },
              {
                "id": "shoes-run-foam-202",
                "name": "Ultra Fly Foam 3",
                "description": "Designed with our softest foam yet, these shoes provide a plush, stable ride for every mile.",
                "price": "150.00€",
                "url": "https://example.com/products/ultra-fly-foam"
              }
            ]
          }
        }
      }
    ]
  }
]

The code to trigger the second step:


    const response = await gemini.models.generateContent({
        model: model,
        contents,
        config: {
            thinkingConfig: { thinkingLevel: ThinkingLevel.LOW },
        }
    });

And the result:

{
  "parts": [
    {
      "text": "To help you find the best match, here are some highly cushioned road running shoes designed for maximum shock absorption:\n\n1.  **CloudStratus Max (169.95€)**: These feature double layers of CloudTec® cushioning. They are designed specifically for road running and are perfect for medium to long distances where you need the most protection from impact.\n2.  **Ultra Fly Foam 3 (150.00€)**: These shoes use a specialized soft foam to provide a \"plush\" and stable ride. They are an excellent choice if you are looking for a very soft feel underfoot during every mile.\n\nWhen trying these on, make sure to wear the socks you plan to run in to ensure the fit is exactly right!"
    }
  ],
  "role": "model"
}

I tried different things without noticeable improvements:

  • using a different role for the function response (‘model’ and ‘function’)
  • different models and different thinking modes
  • prompting : specific instruction, asking to complete the last sentence, even given explicitly

Is there a specific way to structure the history or a specific system instruction that forces Gemini to follow the flow of text ?

Thanks !