Page 124 - CyberSecurityEssentialsEnglish

TABLE 3-6: COMPARING INTERVENTIONS THAT IMPROVE FIT

Intervention The problem it addresses What it does not solve alone

Ambiguity in the task, the A lack of current knowledge,
Prompt
audience, the format, or the exact calculation, or unavailable
engineering
constraints. data.

Selecting and organizing the
Context A poor source, or a fundamental
instructions, files, memory, and
engineering mismatch of the model.
results that reach the model.

Retrieval- Unstable behavior, retrieval
External, private, or updatable
augmented errors, or poor document
knowledge with citation.
generation quality.

Calculation, search, code,
Decisions about permission,
Tool use databases, and the execution of
verification, or workflow design.
actions.

Compliance with a structure or The correctness of the values,
Structured outputs schema for automated the suitability of the content, or

consumption. the quality of the source.

Behavior, style, or a specialized
Fine-tuning or Adding facts that change quickly,
task that recurs over
LoRA or fixing a retrieval problem.
representative examples.

Defining the task, or building a
Sending different tasks to
Model routing shared evaluation across the
different models or tools.
routes.

Organizing multiple steps, A missing capability in a
Workflow design decisions, exceptions, and fundamental component of the

reviews. system.

Retrieval-Augmented Generation: From the Source to the Answer
Retrieval-augmented generation combines the internal memory of the model with an
external, non-parametric memory, meaning one that does not depend on the model’s

parameters and can be updated and queried without retraining the model [6]. The system
begins with trusted documents, cleans them, and adds metadata such as the date, the
issuing body, the version, and the permissions. It then divides them into chunks and converts

them into embeddings or text indexes.

123