Gemini API Safety Settings: Request কেন block হয়

Empty response বা finishReason: SAFETY তিন কারণে হতে পারে: adjustable Harm Category filter, non-adjustable PROHIBITED_CONTENT, অথবা Terms of Service violation। BLOCK_NONE built-in protection bypass করে না।

Safety architecture ও categories

Adjustable safety_settings harm probability (NEGLIGIBLE, LOW, MEDIUM, HIGH) ও threshold ব্যবহার করে। Built-in protection গুরুতর child-safety, PII এবং illegal-harm content-এ প্রযোজ্য। Input block promptFeedback.blockReason, output block candidates[0].finishReason দিয়ে আলাদা করুন; safety_ratings-এ category, probability ও blocked status দেখুন।

Category	Detect করে	সাধারণ trigger
HARASSMENT	identity/protected attribute targeting	quote, role-play, history
HATE_SPEECH	rude/offensive language	profanity
SEXUALLY_EXPLICIT	sexual/obscene content	medical education
DANGEROUS_CONTENT	harmful guidance	security research, weapons

Threshold: OFF, BLOCK_NONE, BLOCK_ONLY_HIGH, BLOCK_MEDIUM_AND_ABOVE, BLOCK_LOW_AND_ABOVE। BLOCK_NONE ratings metadata রাখে। Production-এ settings explicit লিখুন এবং legitimate business-এ আগে narrower threshold পরীক্ষা করুন।

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))

Model version-এর default বদলাতে পারে; explicit settings রাখুন এবং stable GA model ব্যবহার করুন।

PROHIBITED_CONTENT

block_reason: PROHIBITED_CONTENT built-in protection। CSAM, sensitive PII, seriously illegal harm এবং core prohibited category এতে থাকতে পারে। Input review, non-sensitive rephrase, segment করে trigger খোঁজা এবং genuine false positive হলে official support ব্যবহার করুন। Filter bypass বা encoding trick করবেন না।

Debugging ও production fallback

promptFeedback block, candidates না থাকা, finishReason: SAFETY এবং safety_ratings আলাদা log করুন। Model, settings, category probability, request ID ও timestamp record করুন; secret নয়। Fallback হিসেবে safe refusal, narrower rewrite, human review, cache বা alternate workflow রাখুন। Safety refusal-এ endless retry করবেন না।

def safety_summary(response):
    feedback = getattr(response, "prompt_feedback", None)
    if feedback and getattr(feedback, "block_reason", None):
        return {"input_block": str(feedback.block_reason)}
    if not response.candidates:
        return {"status": "no_candidates"}
    c = response.candidates[0]
    return {"finish_reason": c.finish_reason.name,
            "ratings": [(r.category.name, r.probability.name, r.blocked)
                         for r in (c.safety_ratings or [])]}

FAQ

BLOCK_NONE built-in protection সরায় না। blockReason input এবং finishReason output-এর জন্য। Model defaults পরিবর্তনশীল। সব block false positive নয়। Privacy-invasive, deceptive, dangerous বা child-safety request rephrase করে bypass করা যাবে না।

Further Reading

• Error Code Reference