This article explains how to force Google Gemini to output JSON.

---

The Problem

Perhaps like me you are using Gemini and you want to force JSON output. After a first attempting at wrangling you probably keep coming up with either a json text header or a trailing \n line break. Very annoying!

Example of the output:

example__bad_output = """
```json
[
  {"recipe_name": "Chocolate Chip Cookies"},
  {"recipe_name": "Oatmeal Raisin Cookies"},
  {"recipe_name": "Peanut Butter Cookies"},
  {"recipe_name": "Sugar Cookies"},
  {"recipe_name": "Snickerdoodles"},
  {"recipe_name": "Gingerbread Cookies"},
  {"recipe_name": "Shortbread Cookies"},
  {"recipe_name": "Macadamia Nut Cookies"},
  {"recipe_name": "Lemon Cookies"}
]

"""

You might have tried to solve it with some kind of string minipulation but there is actually a better way.


## The solution
So the basic solution is to add the generation_config response_mime_type into your generative AI call like this:

```python
model = GenerativeModel(
    model_name="gemini-1.5-flash",
    generation_config={"response_mime_type": "application/json"},
)

Example Prompt:

prompt = """
    List a few popular cookie recipes using this JSON schema:
    Recipe = {"recipe_name": str}
    Return: list[Recipe]
"""

This will work with both gemini-1.5-flash AND Gemini-pro. If you want to also define a specific output schema this can be done as well but is only supported by Gemini-1.5-pro

response_schema = {
    "type": "ARRAY",
    "items": {
        "type": "OBJECT",
        "properties": {
            "recipe_name": {
                "type": "STRING",
            },
        },
        "required": ["recipe_name"],
    },
}

response = model.generate_content(
    "List a few popular cookie recipes",
    generation_config=GenerationConfig(
        response_mime_type="application/json", response_schema=response_schema
    ),
)

If you want to do this using langchain then you can potentially use the underlying model:

https://python.langchain.com/v0.2/api_reference/core/language_models/langchain_core.language_models.chat_models.BaseChatModel.html#langchain_core.language_models.chat_models.BaseChatModel

Further reading

Code example:

https://github.com/GoogleCloudPlatform/generative-ai/blob/main/gemini/controlled-generation/intro_controlled_generation.ipynb