Gemini API Safety Settings: Requests किन block हुन्छन्

Empty response वा finishReason: SAFETY तीन कारणले आउन सक्छ: adjustable Harm Category filter, non-adjustable PROHIBITED_CONTENT वा Terms of Service violation। BLOCK_NONE ले built-in protection bypass गर्दैन।

Architecture र categories

Adjustable safety_settings ले probability (NEGLIGIBLE, LOW, MEDIUM, HIGH) र threshold प्रयोग गर्छ। Input block promptFeedback.blockReason, output block candidates[0].finishReason बाट छुट्याउनुहोस्; safety_ratings मा category, probability र blocked status हुन्छ। Categories: HARASSMENT, HATE_SPEECH, SEXUALLY_EXPLICIT, DANGEROUS_CONTENT। Thresholds OFF, BLOCK_NONE, BLOCK_ONLY_HIGH, BLOCK_MEDIUM_AND_ABOVE, BLOCK_LOW_AND_ABOVE हुन्। Production मा explicit settings र stable GA model प्रयोग गर्नुहोस्।

from google import genai
from google.genai import types
import os
client = genai.Client(api_key=os.getenv("GEMINI_API_KEY"))
settings = [types.SafetySetting(category=c, threshold=types.HarmBlockThreshold.BLOCK_NONE)
            for c in [types.HarmCategory.HARM_CATEGORY_HARASSMENT,
                      types.HarmCategory.HARM_CATEGORY_HATE_SPEECH,
                      types.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT,
                      types.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT]]
response = client.models.generate_content(
    model="gemini-2.5-flash", contents="your prompt",
    config=types.GenerateContentConfig(safety_settings=settings))

PROHIBITED_CONTENT

यो built-in protection हो: CSAM, sensitive PII, seriously illegal harm र core prohibited categories। Input review, non-sensitive rephrase वा segmentation गर्नुहोस्; वास्तविक false positive भए official support प्रयोग गर्नुहोस्। Filter bypass वा encoding tricks नगर्नुहोस्।

Debugging र fallback

Input block, no candidates, finishReason: SAFETY र safety_ratings अलग log गर्नुहोस्। Model, settings, category probability, request ID र timestamp record गर्नुहोस्, secret होइन। Safe refusal, narrower rewrite, human review, cache वा alternate workflow fallback राख्नुहोस्; safety refusal मा endless retry नगर्नुहोस्।

FAQ

BLOCK_NONE built-in protection हटाउँदैन। blockReason input र finishReason output का लागि हो। Model defaults बदलिन सक्छन्। सबै block false positive होइन। Privacy-invasive, deceptive, dangerous वा child-safety request लाई rephrase गरेर bypass नगर्नुहोस्।

Further Reading

• Error Code Reference