Google Gemini API error - Content generation stopped. Reason: SAFETY

Question

I'm building an android app by using Google Gemini API.

I'm prompting what is 58+78 which is giving correct output. But when I'm prompting what is 2+2, then my app crashes and in Logcat it says : Content generation stopped. Reason: SAFETY.

Complete Logcat :

FATAL EXCEPTION: main
Process: com.example.gemniapi, PID: 22751
com.google.ai.client.generativeai.type.ResponseStoppedException: Content generation stopped. Reason: SAFETY
 at com.google.ai.client.generativeai.GenerativeModel.validate(GenerativeModel.kt:193)
 at com.google.ai.client.generativeai.GenerativeModel.generateContent(GenerativeModel.kt:86)
 at com.google.ai.client.generativeai.GenerativeModel$generateContent$1.invokeSuspend(Unknown Source:15)
 at kotlin.coroutines.jvm.internal.BaseContinuationImpl.resumeWith(ContinuationImpl.kt:33)
 at kotlinx.coroutines.DispatchedTask.run(DispatchedTask.kt:108)
 at kotlinx.coroutines.EventLoop.processUnconfinedEvent(EventLoop.common.kt:68)
 at io.ktor.util.pipeline.SuspendFunctionGun.resumeRootWith(SuspendFunctionGun.kt:135)
 at io.ktor.util.pipeline.SuspendFunctionGun.loop(SuspendFunctionGun.kt:109)
 at io.ktor.util.pipeline.SuspendFunctionGun.access$loop(SuspendFunctionGun.kt:11)
 at io.ktor.util.pipeline.SuspendFunctionGun$continuation$1.resumeWith(SuspendFunctionGun.kt:59)
 at kotlin.coroutines.jvm.internal.BaseContinuationImpl.resumeWith(ContinuationImpl.kt:46)
 at kotlinx.coroutines.DispatchedTask.run(DispatchedTask.kt:108)
 at android.os.Handler.handleCallback(Handler.java:983)
 at android.os.Handler.dispatchMessage(Handler.java:99)
 at android.os.Looper.loopOnce(Looper.java:226)
 at android.os.Looper.loop(Looper.java:328)
 at android.app.ActivityThread.main(ActivityThread.java:9155)
 at java.lang.reflect.Method.invoke(Native Method)
 at com.android.internal.os.RuntimeInit$MethodAndArgsCaller.run(RuntimeInit.java:586)
 at com.android.internal.os.ZygoteInit.main(ZygoteInit.java:1099)
 Suppressed: kotlinx.coroutines.internal.DiagnosticCoroutineContextException: 
  [StandaloneCoroutine{Cancelling}@e794c8a, Dispatchers.Main.immediate]
    

I can't understand what's the SAFETY concern here.

Answer 1

For kotlin for android per question you can use something:

    private val dangerousContent = SafetySetting(HarmCategory.DANGEROUS_CONTENT, BlockThreshold.NONE)
    private val sexuallyExplicit = SafetySetting(HarmCategory.SEXUALLY_EXPLICIT, BlockThreshold.NONE)
    private val hateSpeech = SafetySetting(HarmCategory.HATE_SPEECH, BlockThreshold.NONE)
    private val harassment = SafetySetting(HarmCategory.HARASSMENT, BlockThreshold.NONE)

    private val config = generationConfig {
        temperature = 0.8f
    }

    private val generativeModel = GenerativeModel(
        modelName = "gemini-1.0-pro",
        apiKey = YOUR_API_KEY,
        generationConfig = config,
        safetySettings = listOf(dangerousContent, sexuallyExplicit, hateSpeech, harassment)
    )

You can also play with the temperature:

Temperature = 0: The model will always pick the most likely word or token for each step, making the output highly deterministic (no variation). Useful for factual or exact outputs.

Temperature = 1: The model will generate responses that are more varied and closer to the raw probability distribution, balancing between creativity and coherence.

Temperature > 1: The model will take even more risks in generating responses, often leading to more unpredictable, creative, or experimental results. This could increase the chances of less relevant or unusual responses.

Take always into consideration Ai Principles

Answer 2

The Gemini API has adjustable safety settings. If your prompt is getting this error see why the prompt is blocked and set the filters appropriately.

response = model.generate_content(
    unsafe_prompt,
    safety_settings={
        'HATE': 'BLOCK_NONE',
        'HARASSMENT': 'BLOCK_NONE',
        'SEXUAL' : 'BLOCK_NONE',
        'DANGEROUS' : 'BLOCK_NONE'
    })

More details about safety settings can be found in this Colab Notebook

Answer 3

While using Gemini API, we need to take care of SAFETY Settings. These settings are used to avoid dagenrous/inappropriate being delivered. You can add these like below

from google.generativeai.types import HarmCategory, HarmBlockThreshold

model = genai.GenerativeModel(model_name='gemini-pro-vision')
response = model.generate_content(
    ['Do these look store-bought or homemade?', img],
    safety_settings={
        HarmCategory.HARM_CATEGORY_HATE_SPEECH: HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
        HarmCategory.HARM_CATEGORY_HARASSMENT: HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
    }
)

Ref: To find all the safety settings, please check the documentation https://ai.google.dev/docs/safety_setting_gemini