Results post-processing
Before presenting the raw outputs from GenAI models directly to end users, additional post-processing is often essential to refine and polish results. There are a few common techniques to improve quality, as we will see now.
Filtering inappropriate content – Despite making the best efforts during training, models will sometimes return outputs that are biased, incorrect, or offensive. Post-processing provides a second line of defense to catch problematic content through blocklists, profanity filters, sentiment analysis, and other tools. Flagged results can be discarded or rerouted to human review. This filtration ensures only high-quality content reaches users.
Models such as Google Gemini allow you to define a set of safety settings to set thresholds during generation, allowing you to stop generating content if those thresholds are exceeded. Additionally, it provides a set of safety ratings with your results, allowing you to determine the threshold...
