Prompt pre-processing
Before handing off prompts to generative models, pre-processing can make inputs more usable and potentially improve the quality of the outputs.
When thinking about prompt pre-processing, there are two key dimensions that are affected – security and model usability.
On the security aspect, this is the first opportunity to evaluate the prompts and verify that they align with your responsible AI guardrails. Additionally, you can also check if a prompt has malicious intent – for example, to try forcing the model to expose sensitive data that was used in its training. Putting in place content filters, blocklists, and other defenses at this pre-processing stage is important for ensuring security.
The second dimension is related to optimizing model usability. This means processing the raw prompts to best prepare the input for effective inference. As an example, models are unlikely to accept high-fidelity 192 - kHz audio when probably 8 kHz...
