Week 1 · Day 5/30
RAG Fundamentals
Embeddings, vector databases, chunking strategies, retrieval methods
Celkový progres
17%
🎯 Cieľ dňa
Pochopiť a implementovať RAG pipeline — od chunkingu cez embeddings po vector search a generáciu.
📚 Study Resources
Amir's Blog — RAG Embeddings, Chunking & Retrieval Basics
Jasný, praktický walkthrough troch core RAG komponentov.
articleWeaviate — Chunking Strategies for RAG
In-depth analýza chunking stratégií s benchmarkmi od vector DB makera.
articleFirecrawl — Best Chunking Strategies (2026)
Recursive 512-token splitting: 69% accuracy vs semantic chunking: 54%. Defaults: 256-512 tokens, 10-20% overlap.
benchmarkElephas — 13 Best Embedding Models (2026)
OpenAI text-embedding-3, Voyage 3.5, Nomic V2, Cohere embed-v4. MTEB benchmarks + pricing.
comparisonFirecrawl — Best Vector Databases (2026)
FAISS, Chroma, Pinecone, Qdrant, Weaviate. Chroma pre prototyping, Qdrant pre hybrid.
comparison💡 Key Concepts
Embeddings — Konverzia textu na vysoko-dimenzionálne vektory; cosine similarity, dot product
Chunking Strategies — Fixed-size (simple), recursive (best 69%), semantic (groups by meaning 54%)
Chunk Size Sweet Spot — 256-512 tokenov s 10-20% overlapom ako validované defaults
Vector Databases — In-memory (FAISS) vs managed (Pinecone) vs self-hosted (Qdrant, Weaviate)
RAG Pipeline Flow — Ingestion → Chunking → Embedding → Storage → Query → Retrieval → Generation
🔧 Praktické cvičenie
Buildni document Q&A systém.
- Vezmi 3-5 PDF/markdown dokumentov na tému ktorú poznáš
- Implementuj chunking: recursive text splitter, 512 tokenov, 50 overlap
- Generuj embeddings (Nomic Embed alebo OpenAI small)
- Uložt do ChromaDB (najľahší lokálny setup)
- Query natural language otázkami, vráť top-3 chunks
- Pošli chunks + otázku do LLM. Porovnaj odpovede s/bez RAG