AI Engineering Cheat Sheet

Updated 2026-07-27

Next Topic: AI Evaluation Frameworks and LLM Benchmarking Cheat Sheet

🎯Take a practice test on this topic5 practice tests · 176 questions→

🧠Study flashcards on this topic134 cards · spaced repetition→

AI Engineering is the discipline of building, deploying, and maintaining production-ready applications powered by foundation models — large language models, vision-language models, and multimodal systems. Unlike traditional machine learning, AI engineering focuses on integrating pre-trained models through techniques like prompt engineering, retrieval-augmented generation (RAG), fine-tuning, and increasingly agentic architectures rather than training models from scratch. The field has matured rapidly since 2022, with the 2025–2026 era marked by the shift from prompt engineering to context engineering, the rise of multi-agent frameworks, standardized protocols like MCP and A2A, and the growing importance of evaluation-driven development where reliability, cost-efficiency, and observability in production are as critical as model capability.

What This Cheat Sheet Covers

This topic spans 16 focused tables and 141 indexed concepts, 134 flashcards, 5 practice tests with 176 questions. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Core Prompt and Context Engineering TechniquesTable 2: Retrieval-Augmented Generation (RAG) PatternsTable 3: Data Pipelines and Chunking StrategiesTable 4: Vector Databases and EmbeddingsTable 5: Fine-Tuning MethodsTable 6: Model Optimization and QuantizationTable 7: Model Deployment and ServingTable 8: AI Agent Frameworks and PatternsTable 9: Agent Protocols and InteroperabilityTable 10: Evaluation and TestingTable 11: Observability and MonitoringTable 12: AI Safety and GuardrailsTable 13: Cost Optimization StrategiesTable 14: Advanced Inference TechniquesTable 15: Model Context and Memory ManagementTable 16: Testing and Deployment Practices

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: Core Prompt and Context Engineering Techniques

How you ask a foundation model is most of how well it performs. These range from the bread and butter moves like zero-shot, few-shot, and chain of thought prompting, to reasoning heavy techniques like self-consistency and tree of thought, and finally to context engineering, the 2025 to 2026 shift toward designing the entire information environment a model sees rather than tweaking a single instruction.

Technique	Example	Description
Zero-Shot Prompting	`Translate the following to French: "Hello world"`	Asking the model to perform a task without examples, relying entirely on pre-training knowledge.
Few-Shot Learning	`Q: What is 2+2? A: 4` `Q: What is 5+3? A: 8` `Q: What is 7+1? A:`	• Providing examples in the prompt to teach the model a pattern • typically 3 to 5 examples works well, though the ideal count varies by task and model.
Chain-of-Thought (CoT)	`Let's solve step by step:` `1. First, identify the key variables` `2. Then, calculate...`	• Prompting the model to show reasoning steps • dramatically improves performance on complex reasoning tasks.
Role Prompting	`You are a senior software architect with 15 years experience...`	Assigning a specific persona to shape the model's tone, depth, and perspective.
Self-Consistency	Generate 5 reasoning paths, then vote on most common answer	Running the same prompt multiple times with temperature > 0 and selecting the most frequent answer to reduce errors.
Tree-of-Thought (ToT)	Explore multiple reasoning branches, evaluate each, backtrack if needed	• Extension of CoT that explores multiple reasoning paths simultaneously • useful for strategic planning tasks.
Context Engineering	System prompt + retrieved docs + tool results + memory → single context	• Designing the full information environment an LLM sees at inference • supersedes simple prompt engineering for production systems.
Meta Prompting	`You are a prompt engineer. Generate a better prompt for: {task}`	Using the LLM to improve or generate prompts, creating self-improving prompt pipelines.
Structured Output Prompting	`response_format={"type": "json_schema", "schema": {...}}`	• Forcing model output to conform to a JSON schema • ensures parseable, type-safe results.
Prompt Caching	Cache system prompt processing; only process new user input	• Reusing processed prompt prefixes across requests • up to 90% cost reduction and 85% lower latency (Anthropic).

Table 2: Retrieval-Augmented Generation (RAG) Patterns

RAG is how you give a model knowledge it wasn't trained on without retraining it. The patterns here form a ladder, basic retrieve-and-inject at the bottom, then hybrid search and reranking to sharpen what comes back, up to agentic, self-correcting, and graph-based variants that reason about whether the retrieval was even good enough before generating.

Pattern	Example	Description
Basic RAG	Query → Retrieve docs → Inject into prompt → Generate	• Retrieves relevant documents and injects them into context before generation • addresses knowledge cutoff and reduces hallucination, but cannot eliminate it.

The rest of this section is part of the full cheat sheet

Table 3: Data Pipelines and Chunking Strategies

Before retrieval can work, raw documents have to be parsed and split into pieces small enough to embed but large enough to stay meaningful, and how you cut them shapes everything downstream. Fixed-size splitting is the simple baseline, while semantic and late chunking aim to preserve meaning across boundaries. The pipeline plumbing (metadata, caching, incremental indexing, validation) is what keeps the whole thing fast and trustworthy as your corpus grows.

Technique	Example	Description
Fixed-Size Chunking	Split every 512 tokens with 50-token overlap	• Simplest method • overlap preserves context across boundaries but may split semantic units.

The rest of this section is part of the full cheat sheet

Table 4: Vector Databases and Embeddings

The storage layer that makes semantic search possible. The first rows compare the popular vector databases, from fully managed Pinecone to the Postgres native pgvector, and the tradeoffs between zero ops convenience and keeping vectors next to your relational data. The later rows (HNSW, IVF, product quantization) are the indexing algorithms underneath them, where you tune the balance between speed, recall, and memory.

Component	Example	Description
Pinecone	Cloud-native, serverless scaling, metadata filtering	• Fully managed vector database handling indexing, scaling, and availability automatically • zero-ops.

The rest of this section is part of the full cheat sheet

Table 1: Core Prompt and Context Engineering Techniques

Technique	Example	Description
Zero-Shot Prompting	`Translate the following to French: "Hello world"`	Asking the model to perform a task without examples, relying entirely on pre-training knowledge.
Few-Shot Learning	`Q: What is 2+2? A: 4` `Q: What is 5+3? A: 8` `Q: What is 7+1? A:`	• Providing examples in the prompt to teach the model a pattern • typically 3 to 5 examples works well, though the ideal count varies by task and model.
Chain-of-Thought (CoT)	`Let's solve step by step:` `1. First, identify the key variables` `2. Then, calculate...`	• Prompting the model to show reasoning steps • dramatically improves performance on complex reasoning tasks.
Role Prompting	`You are a senior software architect with 15 years experience...`	Assigning a specific persona to shape the model's tone, depth, and perspective.
Self-Consistency	Generate 5 reasoning paths, then vote on most common answer	Running the same prompt multiple times with temperature > 0 and selecting the most frequent answer to reduce errors.
Tree-of-Thought (ToT)	Explore multiple reasoning branches, evaluate each, backtrack if needed	• Extension of CoT that explores multiple reasoning paths simultaneously • useful for strategic planning tasks.
Context Engineering	System prompt + retrieved docs + tool results + memory → single context	• Designing the full information environment an LLM sees at inference • supersedes simple prompt engineering for production systems.
Meta Prompting	`You are a prompt engineer. Generate a better prompt for: {task}`	Using the LLM to improve or generate prompts, creating self-improving prompt pipelines.
Structured Output Prompting	`response_format={"type": "json_schema", "schema": {...}}`	• Forcing model output to conform to a JSON schema • ensures parseable, type-safe results.
Prompt Caching	Cache system prompt processing; only process new user input	• Reusing processed prompt prefixes across requests • up to 90% cost reduction and 85% lower latency (Anthropic).

Table 2: Retrieval-Augmented Generation (RAG) Patterns

Pattern	Example	Description
Basic RAG	Query → Retrieve docs → Inject into prompt → Generate	• Retrieves relevant documents and injects them into context before generation • addresses knowledge cutoff and reduces hallucination, but cannot eliminate it.

Table 3: Data Pipelines and Chunking Strategies

Technique	Example	Description
Fixed-Size Chunking	Split every 512 tokens with 50-token overlap	• Simplest method • overlap preserves context across boundaries but may split semantic units.

Table 4: Vector Databases and Embeddings

Component	Example	Description
Pinecone	Cloud-native, serverless scaling, metadata filtering	• Fully managed vector database handling indexing, scaling, and availability automatically • zero-ops.