Tier 2: Customize
Go beyond API calls and build AI systems that work with your own data. Learn embeddings, vector databases, RAG pipelines, and agentic architectures that remember, reason, and act.
What You'll Learn
This tier takes you from calling pre-built APIs to building custom AI systems. You'll transform raw text into vector embeddings, store them in purpose-built databases, and wire up retrieval-augmented generation so your models answer from your data—not just their training set. Then you'll build autonomous agents that maintain conversation history, form persistent memories, use tools, query knowledge bases, and coordinate as multi-agent teams.
Part 1: Working with Text
Master embeddings, vector databases, and RAG—the foundation for any AI system that needs to understand and retrieve your own data.
Starting from first principles, you'll learn what embeddings are, how to generate them locally, and how to store and search them at scale. You'll build up to a complete RAG system that answers questions grounded in your documents.
Lesson 1: Embeddings — Understand vectors and embeddings as numeric representations of meaning. Learn dimensionality trade-offs and similarity metrics (cosine, dot product, Euclidean), then generate and compare embeddings with SentenceTransformers.
Lesson 2: Introducing Ollama — Run embedding models locally instead of relying on cloud APIs. Choose between models like bge-m3 and nomic-embed-text, understand latency and memory trade-offs, and swap in Ollama as your embedding backend.
Lesson 3: Vector DBs Basics — Store and search embeddings efficiently with Qdrant. Learn collections, points, payloads, payload filters, and how HNSW indexing enables fast similarity search at scale.
Lesson 4: Adding Traditional Storage — Combine vector search with traditional databases. Understand why vector DBs should only store embeddings and metadata, then link Qdrant semantic search results to full records in SQLite via shared identifiers.
Lesson 5: Chunking & LangChain — Break large documents into embeddable chunks. Use LangChain to load, split, and embed PDFs, then store chunks with unique IDs in Qdrant to avoid duplicates.
Lesson 6: Building a RAG — Put it all together into a Retrieval-Augmented Generation system. Combine embeddings, vector search, chunking, and an LLM to answer questions grounded in your own documents.
Lesson 7: Multi-Dimensional Data Visualization — Visualize high-dimensional embeddings in 2D using UMAP and t-SNE. Learn to read clusters, distances, and outliers from dimensionality-reduced plots without fooling yourself.
Part 2: Agentic AI
Build autonomous AI agents with agno—from stateless bots to multi-agent teams with memory, tools, and knowledge bases.
This section moves from data retrieval to autonomous behavior. You'll set up infrastructure (PostgreSQL, Qdrant, Arize Phoenix), then progressively add capabilities: conversation history, persistent memory, tool use, RAG knowledge, and multi-agent coordination.
Lesson 8: Setup — Configure the development stack: Docker, PostgreSQL for storage, Qdrant for vectors, and Arize Phoenix for observability. Understand why agents need more than just an API key.
Lesson 9: Basic Agent — Build the simplest agent: system prompt + user input → LLM → output. Understand the agent loop, stateless design, and what's missing compared to production chat systems.
Lesson 10: History — Add session-based conversation history with PostgreSQL. Support ephemeral and persistent sessions so your agent remembers what was said earlier in a conversation.
Lesson 11: Memory — Give your agent cross-session memory. The LLM automatically identifies and stores memorable facts per user, persisting them across sessions and script restarts.
Lesson 12: Tools — Extend agents with real-world capabilities. Define native Python tools and connect external services via MCP. Understand the tool-calling flow—how the LLM decides when and what to call.
Lesson 13: Knowledge — Ground agent responses in your documents using RAG. Convert PDFs and URLs into chunked, searchable vector embeddings with Qdrant and compare fixed-size vs semantic chunking strategies.
Lesson 14: Teams — Combine specialized agents into coordinated teams. Build specialist agents with focused roles, wire up a team leader for routing, and manage context sharing between agents.
Why This Matters
Part 1 gives you the complete pipeline from raw text to semantic search—the backbone of any AI system that works with custom data. Part 2 takes those retrieval capabilities and embeds them into autonomous agents that remember users, use tools, and collaborate.
Master Tier 2 and you can build AI systems that go far beyond simple prompt-and-response: systems that understand your data, learn from interactions, and take actions in the real world.