LLM Zoomcamp

DataTalks.Club · 2026 · Building LLM applications from scratch

7-module course covering RAG, vector search, orchestration, evaluation, monitoring, and best practices. Code for all modules lives in the GitHub repo.

Modules

5 of 7 modules completed

Completed 01

Agentic RAG

Built a RAG system from scratch — search, prompt engineering, chunking — then made it agentic with function calling and an LLM-driven loop.

minsearchchunkingfunction callingagentic loop
Completed 02

Vector Search

Semantic search with embeddings — moving beyond keyword matching to meaning-based retrieval. Built vector search by hand with numpy, then with minsearch, sqlitesearch, and PGVector; combined keyword and vector search with hybrid RRF fusion.

embeddingsONNXchunkinghybrid search (RRF)
Completed 03

Orchestration

AI orchestration with Kestra — context engineering, AI Copilot, RAG (static + web search), autonomous agents, and multi-agent delegation, all wrapped in versioned, observable workflows.

KestraRAGAIAgentmulti-agent
Completed 04

Evaluation

Measuring whether a RAG system actually works — generating ground truth with an LLM, scoring retrieval with Hit Rate and MRR, comparing text vs. vector search, and fusing both with Reciprocal Rank Fusion.

ground truthHit RateMRRRRF
Completed 05

Monitoring

Moving from offline evaluation to production observability — instrumenting a RAG pipeline with OpenTelemetry traces and span attributes, persisting them to SQLite, and querying where time, tokens, and cost actually go.

OpenTelemetrytracesSQLiteGrafana
Upcoming 06

Best Practices

Hybrid search, guardrails, and production patterns for reliable LLM systems.

hybrid searchguardrails
Upcoming 07

Project

End-to-end LLM application — putting everything together.

capstone