GCP GenAI Leader cheat sheet

Generative AI creates new content (text, images, code, audio, video) by predicting probable next tokens; discriminative AI instead assigns labels or scores to existing data.

Tokenization rule of thumb: about 4 characters of English text equals 1 token; prompts, history, and outputs all consume tokens and are billed per token volume.

Context window = input prompt + conversation history + generated output; anything beyond the window is truncated or rejected, so long chats need summarization or retrieval.

Temperature and top-p control sampling: near-zero temperature suits extraction, classification, and deterministic workflows; higher values (roughly 0.7–1.0) suit brainstorming and creative variants.

Hallucinations occur because models optimize for plausible text, not verified truth; mitigate with grounding, tuning on domain data, and citation requirements before trusting output.

Embeddings map text, images, or audio to vectors; similarity between vectors (for example cosine distance) powers semantic search, clustering, deduplication, and RAG retrieval.

Multimodal models accept mixed inputs such as images, PDFs, audio, and video alongside text, removing the need for separate OCR or transcription pipelines.

Model-family decision rule: larger reasoning-oriented models for complex analysis; Flash-class or smaller models for high-volume, low-latency, cost-sensitive tasks.

Diffusion models generate images by iteratively denoising random noise, while transformer-based LLMs generate text one token at a time conditioned on prior context.

Vertex AI is the unified platform: Model Garden, Vertex AI Studio, tuning, evaluation, Vector Search, pipelines, and prediction endpoints sit behind one console and API.

Model Garden catalogs Google first-party models (Gemini, Imagen, Veo, Chirp, Embeddings), third-party partner models, and open models that can be deployed to managed endpoints.

Gemini tiers follow a cost/latency rule: Pro-class for complex reasoning and agentic tasks, Flash-class for high-throughput low-latency workloads, and lightweight variants for constrained environments.

Vertex AI Studio is the prototyping surface for prompt design and model comparison; production traffic moves to Vertex AI API endpoints authenticated with IAM.

Grounding options differ by source: Grounding with Google Search adds fresh web results plus citations; grounding with your own content uses Vertex AI Search datastores built from enterprise documents.

Task-to-model mapping: Imagen for text-to-image, Veo for text-to-video, Chirp for speech-to-text, and the Embeddings API for vector representations of text or images.

Vertex AI Agent Builder and Conversational Agents combine instructions, tools, and enterprise connectors to build task-oriented agents and virtual assistants.

Gemini for Google Cloud (formerly Duet AI) assists in Cloud console, BigQuery, and code editors, while Gemini for Google Workspace spans Docs, Gmail, and Meet.

Vertex AI Model Evaluation and the Gen AI evaluation service score outputs for quality, safety, and groundedness using automatic metrics and model-based raters.

Decision rule: missing private knowledge → grounding (RAG); wrong style, format, or task behavior → tuning; vague or inconsistent results → better prompting first.

Prompt anatomy: task instruction + context/data + examples + output format + constraints; the system instruction carries persistent role and safety rules across turns.

Few-shot prompting with roughly 3–5 well-chosen examples is the cheapest fix for format and tone errors and requires no retraining.

Chain-of-thought prompting ('reason step by step') improves multi-step math, logic, and planning tasks, but adds latency and is unnecessary for simple extraction.

RAG pipeline steps: chunk documents → embed → index in a vector store → retrieve top-k matches per query → inject chunks into the prompt → generate an answer with citations.

Supervised fine-tuning needs labeled input-output pairs, typically hundreds to thousands, and is chosen when behavior must be consistent beyond what prompting achieves.

Adapter (parameter-efficient) tuning trains a small set of extra weights rather than the full model, lowering cost and preserving the base model.

Distillation transfers a larger model's behavior into a smaller, cheaper model to cut latency and per-token cost in production.

Constrain output explicitly (JSON schema, allowed values, maximum length) and validate responses programmatically before downstream systems consume them.

Google's AI Principles commit to socially beneficial use, avoiding unfair bias, safety testing, accountability, privacy by design, and scientific excellence; high-risk use cases get formal review.

Data controls: Google Cloud does not use customer prompts or data to train foundation models without permission; VPC Service Controls, CMEK, and IAM restrict access and movement.

SAIF (Secure AI Framework) addresses threats such as model theft, data poisoning, prompt injection, and unsafe output; Model Armor screens prompts and responses for these risks.

Responsible deployment gates: risk classification, safety filters and thresholds, human-in-the-loop for consequential decisions, and ongoing monitoring for drift and abuse.

Adoption sequence: select a bounded high-value use case → pilot with measurable KPIs → build enablement and a center of excellence → scale with governance in place.

Change management levers that correlate with successful rollout: executive sponsorship, an internal champions network, AI literacy training, and a clear acceptable-use policy.

Cost governance: track token consumption per feature, route high-volume simple tasks to Flash-class models, cache repeated prompts, and cap max output tokens.

Build-versus-buy rule: use prebuilt Google models and agents for common tasks, and invest in tuning or custom models only where differentiation genuinely requires it.