Core Concepts · Review
The Full Playbook
12 concepts. Read, revisit, then hit Quiz when you're ready.
▸ Foundations
Concept 01✓ Covered
LLM — Large Language Model
The engine underneath everything.

An LLM is software trained to predict text. Feed it words, it guesses what comes next — billions of times, on billions of sentences, until the guesses get eerily good.

Metaphor
That autocomplete on your phone keyboard — but trained on essentially all of human writing and scaled up massively.
Training happens once, takes months, requires enormous compute. After that, the model is frozen.
The model doesn't look things up — it pattern-matches from training data.
This is why it can be wrong: if something wasn't in the training data, it can't know it — and may hallucinate instead.
Concept 02✓ Covered
Tokens & Context Window
How LLMs actually read — and how much they can hold.

LLMs don't read words — they read tokens: chunks roughly ¾ of a word each. "Unbelievable!" = 4 tokens. 1,000 tokens ≈ 750 words ≈ 3 pages.

Metaphor
The context window is a whiteboard. Everything — your question, the conversation history, any documents — has to fit. When it's full, older content gets erased.
Every LLM has a token limit (the context window). Claude's is large; GPT-4 varies by version.
Once the window fills, the LLM "forgets" what was said earlier.
You can't dump an entire knowledge base into one prompt — this is the core problem RAG solves.
Concept 03✓ Covered
RAG — Retrieval-Augmented Generation
The open-book test approach.

RAG is a pattern: before the LLM answers a question, a retrieval system searches your data, grabs the most relevant chunks, and stuffs them into the prompt. The LLM then generates an answer grounded in that real data.

Metaphor
It's like an open-book exam. The student (LLM) doesn't need to memorize everything — a librarian (retriever) hands them exactly the right pages before they answer.
Why not fine-tuning? RAG is faster to set up, doesn't require retraining, and your data stays current.
Quality depends on the retriever. If the wrong chunks are fetched, the LLM generates a confidently wrong answer.
Healthcare angle: RAG lets a health system ground answers in its own formulary or care protocols — without exposing PHI to model training.
The retrieval step usually relies on embeddings (Concept 04) to find semantically similar content.
Concept 04✓ Covered
Embeddings
How meaning becomes math.

An embedding is a list of numbers (a vector) that represents the meaning of a piece of text. Similar meanings produce similar vectors. "Heart attack" and "myocardial infarction" land close together even though they share zero words.

Metaphor
Imagine every piece of text gets a GPS coordinate in "meaning space." Texts about the same topic cluster together, even if they use completely different words.
Embeddings power the "retrieval" step in RAG. Embed your knowledge base, embed the question, find the nearest neighbors.
Stored in a vector database — a specialized DB optimized for nearest-neighbor search.
Also used for recommendation engines, semantic search, clustering, and anomaly detection.
Healthcare angle: Semantic search across clinical notes, matching patients to clinical trials, surfacing similar prior-auth cases.
▸ Working with Models
Concept 05✓ Covered
Fine-Tuning
Teaching the model new tricks — permanently.

Fine-tuning takes a pre-trained LLM and continues training it on your specific data. The model's weights actually change — your knowledge gets baked into the parameters.

Metaphor
If the base LLM graduated with a general degree, fine-tuning is sending it to a specialized residency program.
RAG vs. Fine-Tuning: RAG = open-book exam (always current). Fine-tuning = memorizing the textbook (frozen at training time). Most enterprise use cases start with RAG.
Fine-tuning is best for teaching style, tone, or behavior — not for injecting factual knowledge that changes often.
It's expensive: curated training data, GPU time, evaluation. Risk of overfitting or "catastrophic forgetting."
Customer objection: "Can't we just fine-tune it on all our data?" — Usually no. RAG handles dynamic knowledge; fine-tuning handles behavioral patterns.
Concept 06✓ Covered
Prompt Engineering
The art of asking the right way.

Prompt engineering is the practice of structuring your input to an LLM to get the best possible output. The model is the same — the prompt is the lever you pull.

Metaphor
You're briefing a brilliant contractor. Vague brief = vague result. Specific brief with examples, constraints, and a clear deliverable = great work.
Key techniques: Role assignment ("You are a clinical pharmacist…"), few-shot examples, chain-of-thought, output format specification.
System prompts set persistent instructions. User prompts carry the specific request.
Prompt engineering is often the highest-ROI lever. Before fine-tuning or RAG, try better prompts — it's free and instant.
Demo tip: Show a bad prompt vs. a good prompt side by side. The quality delta lands immediately with buyers.
Concept 07✓ Covered
Hallucination & Grounding
Why models make stuff up — and how to stop it.

Hallucination is when an LLM generates text that sounds confident but is factually wrong or entirely fabricated. It's not a bug — it's a consequence of how prediction works.

Metaphor
A student who never says "I don't know." Grounding is giving them a cheat sheet and telling them: only answer from what's on this page.
Why it happens: LLMs optimize for plausible next tokens, not truth.
Grounding strategies: RAG, citations, constrained output, temperature reduction.
Healthcare angle: #1 concern in regulated industries. "What if it hallucinates a drug interaction?" Grounding + guardrails + human-in-the-loop.
Demo tip: Show the same question with and without grounding context. The contrast sells RAG.
▸ Agents & Orchestration
Concept 08✓ Covered
Agents
LLMs that can do things, not just talk.

An agent is an LLM given the ability to take actions: call APIs, query databases, trigger workflows. It receives a goal, reasons about it, picks tools, executes, observes results, and iterates.

Metaphor
A regular LLM is a consultant who gives advice. An agent is a consultant with admin credentials — they can actually go do the work.
Core loop: Think → Act → Observe → Repeat.
Tools are the differentiator. Without tools, it's just a chatbot.
Guardrails matter more. Real actions = real consequences. Human-in-the-loop for high-stakes actions.
Adobe angle: AI Assistant in AEP is an agent — it queries data, builds segments, surfaces insights.
Concept 09✓ Covered
MCP — Model Context Protocol
How agents get their tools.

MCP is an open standard (created by Anthropic) that defines how an AI model discovers and connects to external tools and data sources. Universal plug-and-play for AI.

Metaphor
USB-C for AI. Before USB-C, every device had a different charger. MCP is the universal connector.
How it works: An MCP server exposes tools with descriptions. The model reads them, decides when to call them, sends structured requests.
Why it matters: Without MCP, every AI integration is bespoke. With MCP, one server = every AI client can use it.
Adobe context: Adobe has MCP servers for AEP, Express, Analytics, AEM, Firefly, and more.
Competitive context: MCP is the most widely adopted open standard as of 2025.
Concept 10✓ Covered
Agent-to-Agent (A2A)
When agents stop working alone and start collaborating.

A2A is a protocol (led by Google) that allows AI agents from different vendors to discover, communicate with, and delegate tasks to each other.

Metaphor
A hospital's referral system. The ER doctor triages, then hands off to specialists. A2A is the communication protocol between them.
A2A vs. MCP: MCP connects agents to tools. A2A connects agents to other agents. Complementary layers.
Agent Cards: Each agent publishes a description of what it can do. Other agents discover it, negotiate tasks, exchange messages.
Enterprise angle: Large orgs have agents across vendors. A2A lets them collaborate without point-to-point integrations.
Healthcare scenario: Scheduling agent → benefits-verification agent → prior-auth agent. Seamless patient experience.
▸ Enterprise & Safety
Concept 11✓ Covered
Multi-Modal AI
Beyond text — seeing, hearing, generating across formats.

Multi-modal AI can process and generate more than just text — images, audio, video, documents — within the same conversation.

Metaphor
Early AI was a radio. Multi-modal AI is a person in a meeting: reads slides, hears the speaker, watches demos, takes notes, sketches diagrams. All at once.
Input: Text, images, PDFs, audio, video, structured data.
Output: Text, code, images (Firefly), structured data. Frontier: real-time video and voice.
Adobe angle: Firefly for images, Premiere for video AI, Acrobat for documents — all multi-modal, integrated.
Healthcare angle: Medical imaging analysis, voice notes → EHR entries, patient education from clinical text.
Concept 12✓ Covered
Guardrails & Responsible AI
The safety rails that make enterprise AI deployable.

Guardrails are policies, filters, and patterns that prevent AI from producing harmful, biased, or non-compliant outputs. Responsible AI is the broader discipline of fairness, transparency, and accountability.

Metaphor
Highway median. The car (LLM) is powerful, but without barriers it can cross into oncoming traffic. Guardrails keep it in its lane.
Types: Input filtering, output filtering, topic boundaries, PII detection, human-in-the-loop.
Bias: Models amplify training data biases. In healthcare, this means worse outcomes for underrepresented populations.
Regulatory: EU AI Act, HIPAA, state-level AI laws. Auditability and explainability are table stakes.
Adobe angle: AI Ethics principles, Content Credentials, Firefly's commercially safe training data.