Advanced Prompting · Tool Calling · RAG — from theory to running code
Everything we built yesterday in 60 seconds
Symbolic AI → ML → DL → Transformers → LLMs
Gradient descent, backpropagation, loss landscapes
BPE subword pieces, ~4 chars ≈ 1 token
Words as vectors: King − Man + Woman ≈ Queen
Q·K→weights, V→output. Every token sees every other.
Pre-training → SFT → RLHF. Next-token prediction at scale.
Good prompting = 10× productivity. Using an AI agent well is a skill in itself.
A single well-structured conversation can produce a complete, multi-file study guide with:
The key? Context + Structure + Iteration. Give the agent your syllabus, question papers, and textbook extracts. Tell it the output format you want. Let it build iteratively.
Adding "Let's think step by step" fundamentally changes how the model reasons
Waiting...
Waiting...
Assigning a persona fundamentally changes tone, expertise, and perspective
"Explain quantum computing"
"Quantum computing leverages quantum mechanical phenomena such as superposition and entanglement to process information..."
"Imagine you have a magic coin that can be heads AND tails at the same time..."
"Arrr! Imagine ye have a treasure map where X marks ALL the spots at once, savvy?"
During training, the model learned patterns from text written by people in different roles. When you set a role, you're activating the specific subset of knowledge and style associated with that persona in the model's weights.
👇 Try it live in the next section
Paste an image URL → AI creates a persona → Chat with the character
Character image will appear here
Character details will appear after analysis...
LLMs can't browse the web, run code, or query databases — but they can decide to call tools that do
User
LLM
Decide: respond
or call a tool?
Tool Functions
get_weather(city="London")
Query a weather API
search_database(query="revenue Q4")
Search internal data
run_python(code="2**32")
Execute code in a sandbox
send_email(to="team@co.com", body="...")
Trigger real-world actions
The LLM doesn't execute these functions — it outputs a structured JSON describing what function to call and with what arguments. Your application code executes the function and feeds the result back.
A text adventure powered by tool calling — watch the LLM choose functions in real-time
Dungeon
Entrance
Library
Treasure
You awaken in a dusty entrance hall. Cobwebs cling to cold stone walls. A rusty torch flickers nearby. Something tells you there's treasure to be found — and a way out.
Type a command like "look around", "go east", or "pick up the torch"
LLMs have limits. RAG extends them.
LLMs are frozen in time. GPT-4o's training data ends at a specific date. It doesn't know what happened yesterday.
LLMs confidently generate plausible-sounding but entirely made-up facts. There's no "I don't know" default.
Your company's internal docs, proprietary data, and domain knowledge aren't in the training data.
RAG = Retrieval + Augmented + Generation
Don't fine-tune the model. Give it the right context at inference time.
Upload a document, watch it get chunked and embedded, then query it
Upload a document, then ask a question...
From simple retrieval to advanced multi-step pipelines
Too small = missing context. Too large = diluted relevance. 200-500 tokens is a good start.
Overlapping chunks ensure information at boundaries isn't lost. 10-20% overlap is common.
OpenAI, Cohere, or open-source (Sentence-BERT). Match your domain and cost.
Pinecone, Weaviate, Chroma, pgvector, Qdrant. Pick based on scale and hosting needs.
Retrieved chunks ranked by embedding similarity may not be the most useful. A re-ranker (Cohere, cross-encoder) improves precision.
Measure: retrieval recall, answer correctness, faithfulness (no hallucination), and latency.