AI Learning Roadmap — Part 2

Intermediate Level

1. Model Context Windows

The amount of text an LLM can “see” at once. Think of it as the model’s short-term memory.

Example:

Model A: 2,000 token window → uploading a 5,000-token document means it only sees the first 2,000 tokens → incomplete answers.

Model B: 100,000 token window → sees everything → complete answers.

Current models (Jan 2026):

Claude 3.5 Sonnet: 200,000 tokens
GPT-4: 128,000 tokens
Older models: 2,000–4,000 tokens

Why it matters:

Large documents need large windows
Bigger windows cost more but improve accuracy
Essential to know before building systems

2. Embeddings and Semantic Search

Converts text into numbers representing meaning. Similar texts → similar numbers; different texts → different numbers.

Example:

"I love cats" → [0.2, 0.5, -0.1, ...]
"I like cats" → [0.21, 0.49, -0.09, ...] (similar)
"I hate snakes" → [0.9, -0.2, 0.5, ...] (different)

Why numbers? Computers can quickly compare numbers to find similarity.

Semantic search:

Keyword search: only finds exact words
Semantic search: finds documents with similar meaning even if words differ

Example

User searches your product documentation for: “How do I reset my password?”

Old way (keyword search):

Only finds documents with exact words “reset” and “password”

Misses document titled “Account Recovery Guide”

New way (semantic search with embeddings):

Converts “reset password” to numbers

Compares those numbers against all documents

Finds “Account Recovery Guide” because it has similar meaning

Works even though exact words don’t match

Vector database

A special database that stores embeddings and finds similar ones fast.

Example

10,000 customer reviews → converted to embeddings → stored

User asks: “Which reviews mention good customer service?”

You convert that question to an embedding.

Vector Database quickly finds the 50 reviews with similar embeddings. Returns them in milliseconds.

3. Retrieval-Augmented Generation (RAG)

A technique to make AI answer questions about YOUR data without retraining the model. It works in 3 steps: Find relevant documents, Give them to the AI, AI answers based on those documents.

Problem it solves:

LLMs trained on public internet data (not your private company data)
You can’t retrain Claude on your data (too expensive, too slow)
Solution: RAG — dynamically feed your data to Claude

Example:

Question: “How many vacation days do I get?”

Without RAG: AI gives a generic answer -> answers “It varies by company, typically 15–20 days…”

With RAG: AI reads your policy → answers “25 days per year”

Why it matters:

Cheaper and faster than retraining
Ensures AI answers are accurate for your data

4. Hallucinations

Hallucination occurs when AI confidently says something false. It’s not lying — it’s predicting plausible text based on patterns it learned.

Examples:

User: “What did Einstein say about quantum mechanics?”

AI: “Einstein once said, ‘The universe operates on quantum principles that transcend classical physics.’”

Reality: Einstein never said exactly that. The AI made up a quote that sounds like something he might say.

How to reduce hallucinations:

Use RAG to provide real data
Ask AI to cite sources
Validate outputs
Lower temperature to reduce creativity

Why it matters: You need to know this exists so you design systems to prevent it. Users trust AI, so false information spreads easily.

5. Fine-Tuning

Training an AI model on your specific data to make it better at your specific task. You start with a base model and customize it.

Simple analogy:

Base model: A student who studied general knowledge

Fine-tuning: Intensive study in just one subject (making them an expert in that subject)

When to use:

Large dataset (500+ examples)
Need specific behavior or tone
Consistency across responses
Reduce token costs over time (fine-tuned model is smaller/faster)

Why it matters: Fine-tuning makes AI applications more useful for specialized tasks without retraining the entire model