810-110: Cisco AI Technical PractitionerDemo

Prompt chaining is the technique of decomposing a complex task into a sequence of smaller, linked prompts, where the output generated by one prompt is used as the input to the next prompt in the sequence. This allows a large or multi-step problem to be solved incrementally, with each prompt handling a discrete subtask and passing its result forward. This differs from few-shot prompting (providing example input/output pairs within a single prompt), meta-prompting (using prompts to generate or optimize other prompts), and chain-of-thought prompting (eliciting step-by-step reasoning within a single prompt/response rather than across multiple sequential prompts).

After establishing a high-level data summary during the orient step, feedback patterns should be visualized to expose distributions, trends, outliers, and differences among customer segments. Correlation and hypothesis development follow this exploratory visualization.

Cloud-hosted API services run on infrastructure managed by the provider, which can automatically and rapidly scale compute resources up or down in response to fluctuating inference workloads. This elastic scalability lets applications handle unpredictable spikes in demand without requiring the customer to provision, maintain, or over-provision physical hardware, which is a key advantage over locally hosted models that are limited by fixed on-premises capacity. Per-token billing means operational costs scale with usage rather than being eliminated, data residency and privacy control are typically stronger with local hosting, and network latency for cloud APIs varies by region and is not guaranteed to be lower for every user.

When a tool is defined for an AI agent, its JSON schema (such as Claude's input_schema or the parameters object in OpenAI-style function definitions) specifies the tool's expected properties, their data types (string, number, boolean, array, enum, etc.), and which properties are mandatory via a required array. The model uses this schema to construct a structurally valid set of arguments when it decides to call the tool, and the calling application can validate the returned input against the same schema before executing the tool. This ensures the agent supplies well-formed, type-correct, and complete parameters rather than free-form text, eliminating syntax and structural errors in tool calls.

Reformatting data from one structure or encoding (CSV) into another (a Markdown table) while preserving its content and meaning is an example of data transformation — the process of converting data from one format or structure into another to make it usable for a particular purpose, such as presenting it in a report. This differs from classification (assigning categories or sensitivity labels to data), filtering (selecting or excluding specific records), and quantization (reducing numeric precision or continuous values into discrete levels), none of which describe a simple format conversion like CSV-to-Markdown-table.

Providing background information and technical constraints in a prompt is an application of the principle of describing the situation — giving the AI model relevant contextual details (such as project background, environment, and limitations) so it can generate a more accurate and relevant response. This differs from defining a persona (assigning the AI a specific role or character), using few-shot prompting (supplying example inputs and outputs to guide the response format), and stating the purpose (clarifying the specific goal or desired outcome of the prompt).

Cloud LLM API providers price usage based on the number of tokens consumed by a request, counting both input tokens (the prompt and any context) and output tokens (the generated response). Since tokenization breaks text into these billable units, the total token count for a request directly determines its cost — this is the standard per-request billing model used by providers such as OpenAI and Anthropic, rather than a flat subscription fee, a simple per-call charge independent of size, or direct GPU time billing.

In high-stakes domains, AI systems can produce errors, exhibit bias, or misjudge edge cases not well represented in training data. A Human-in-the-Loop requirement ensures that a qualified person reviews, validates, or can override an AI system's output before it is acted upon, which is critical for confirming the accuracy and appropriateness of decisions that could seriously affect patient health or financial outcomes. This human oversight mechanism is a core AI governance control specifically aimed at ensuring the validity and trustworthiness of AI-driven decisions, not at data protection, access control, or improving processing speed.

A vector database is designed to store and index vector embeddings — numerical representations of data such as text, images, or audio that capture semantic meaning. In a RAG pipeline, source documents are chunked and converted into embeddings, which are stored in the vector database. At query time, the user's input is also embedded, and the database performs a similarity search (e.g., cosine similarity or nearest-neighbor search) to retrieve the most semantically relevant chunks, which are then passed to the language model as additional context. This is distinct from managing the context window (an LLM inference concern), updating model weights (a training/fine-tuning process), or allocating compute resources (an infrastructure orchestration task).

This is an example of a direct prompt injection attack, in which an attacker submits input directly to a large language model (LLM) application in an attempt to override its original instructions or system prompt and manipulate it into performing unintended actions, such as revealing sensitive information. Because the malicious instruction is entered directly into the chatbot's user input field—rather than being hidden in external content the model later processes (which would be indirect injection)—it is classified as a direct injection attack. This differs from insecure supply chain issues (compromised dependencies or third-party components), model drift (natural degradation of model accuracy over time due to changing data), and data poisoning (corrupting training data to bias model behavior).

The requirements phase of the SDLC focuses on eliciting, analyzing, and formally documenting stakeholder needs, often expressed as user stories or requirement specifications. AI tools applied in this phase can scan stakeholder documentation to detect contradictions, ambiguities, or gaps, and can automatically draft structured user stories from that input, supporting requirements elicitation and analysis. This is distinct from maintenance (post-deployment support and fixes), implementation (writing code), and testing (verifying that the built system meets requirements).

MCP defines a client-server architecture where an MCP host application uses an MCP client to connect to MCP servers. Each MCP server exposes capabilities—tools, resources, and prompts—that an AI model can discover and invoke through the standardized protocol. The server itself does not run model inference (that's the LLM's job), does not provide a user interface (that's the host application), and does not act as a router between AI providers. Its defined role is to make external functionality and data available to the model in a consistent, standardized way.

Generative AI models respond to a single prompt and produce one output in a single pass (single-turn generation), such as text, images, or code, based on patterns learned from training data. Agentic AI extends this by using reasoning, planning, memory, and tool integration to autonomously execute multi-step workflows — breaking a goal into subtasks, taking actions, evaluating results, and iterating in a loop until the objective is achieved, without needing a new prompt for every step. This capacity for multi-step, autonomous reasoning and action is the key differentiator between agentic AI and traditional generative AI.

A coordinator agent operates as the central orchestration layer in a multi-agent system: it decomposes an overall objective into subtasks, delegates each subtask to the specialized agent best equipped to handle it, tracks execution and handles failures, and then synthesizes or aggregates the individual agents' outputs into a single coherent final result. This orchestration-and-aggregation function is what distinguishes a coordinator from the specialized worker agents themselves, and it is a core pattern used in frameworks such as LangGraph, CrewAI, and Microsoft AutoGen.

Diffusion models work by first progressively adding random noise to training images (the forward process) and then training a neural network to reverse this process (the reverse process), gradually removing noise step-by-step to reconstruct or generate a new, high-fidelity image from pure noise. This iterative denoising mechanism is what enables diffusion models to produce highly detailed and realistic images, and it underlies popular image-generation systems such as Stable Diffusion and DALL-E. GANs instead rely on an adversarial generator-discriminator training setup, VAEs use an encoder-decoder architecture with a probabilistic latent space, and LLMs are designed for text generation rather than image synthesis.

Refactoring involves restructuring existing code to improve its internal quality without altering its external behavior. AI-powered tools support this process by analyzing codebases to detect design flaws, code smells, and inefficiencies, then recommending or applying improvements to the code's structure and design. This is distinct from tasks like creating database schemas, integrating APIs, or generating documentation, which are separate development activities not directly tied to the refactoring process.

An approval gate is a control point inserted into an automated or agentic workflow that pauses execution and requires explicit human sign-off before a significant or potentially risky action is carried out. In this case, the AI agent is permitted to autonomously perform the lower-risk task (drafting the email) but the higher-risk, consequential action (actually sending it to the customer) is withheld until a human reviews and authorizes it. This human-in-the-loop checkpoint is a standard responsible-AI and AI-agent governance practice used to maintain oversight and reduce the risk of unintended or harmful autonomous actions, distinguishing it from drift monitoring (tracking model performance degradation), data lineage (tracking data origin/transformation history), and injection mitigation (defending against prompt injection attacks).

Fairness in responsible AI focuses on ensuring that AI systems do not produce biased or discriminatory outcomes and that they treat all individuals and groups equitably. Auditing model outputs for bias is a direct, concrete way to detect and correct disparate or unfair treatment of different demographic groups, which is the core objective of the fairness principle. Other actions—redacting PII (privacy), hosting in carbon-neutral data centers (sustainability), and requiring AUP sign-off (governance/accountability)—address different responsible AI principles, not fairness specifically.

Self-consistency is a prompting strategy in which the same prompt is executed multiple times, and the resulting outputs are compared (often using majority voting) to check whether the model consistently produces the same or similar answer. This helps verify the reliability and stability of the model's output rather than affecting speed, context window limits, or creativity.