Page 128 - CyberSecurityEssentialsEnglish
P. 128

model does not hold to a consistent classification style despite many examples, tuning may
         be appropriate. A system can use retrieval, LoRA, and good prompting together.

         Structured Outputs and Model Routing
         Asking for JSON is a formatting instruction, but it does not guarantee that the structure will

         be valid. In systems that support structured outputs or schema validation, the fields, types,
         and required values can be specified and checked automatically before the result is passed
         on. Even so, the meaning of the values must be verified; a correct structure does not mean

         correct content.

         Model routing sends each task to a model or tool according to its characteristics: a small fast
         model for routine classification, a stronger model for composite analysis, a search engine for
         current facts, and a code tool for calculation. This raises efficiency, but it requires routing

         rules, shared measures, and a fallback path when the classification fails.


           ● Remember: Section summary
           • Key idea: The right intervention follows the cause of the failure.

           • Practical decision: Test the smallest change that addresses the problem, then measure
           its effect.
           • Mistake to avoid: Using fine-tuning to add current information, or using retrieval to
           repair an undefined task.


                                                           Video: What Is Retrieval-Augmented

                                                           Generation?
                                                           A visual explanation of connecting a model to
                                                           a current, citable content store instead of

                                                           relying only on what it learned during
                                                           training.

                                                           Duration: about 7 minutes  |  Level: Beginner

                                                           Learning objective: Distinguish the internal
                                                           memory of a model from an updatable
                                                           external source.

                                                           Source: IBM Technology [24].
                                                           https://www.youtube.com/watch?v=T-
                                                           D1OfcDW1M

                                                           Verified: June 6, 2026











        127
   123   124   125   126   127   128   129   130   131   132   133