This is the goal of the whole roadmap: build things with language models. Prompt them reliably, turn text into searchable embeddings, ground answers in your own documents with RAG (that cites its sources), and give an LLM tools so it can act — while keeping the judgment to know when a plain workflow beats an agent. Free-tier throughout: Ollama locally, free API tiers, Colab GPUs.
Ship two things and you've arrived: (1) "chat with your documents" RAG with numbered citations and a RAGAS eval, deployed (Streamlit/Gradio); (2) a tool-using agent that uses your RAG retriever as one of its tools. Pick your domain — a security analyst (RAG over CVE/ATT&CK + a triage agent) or an investing research assistant (RAG over filings). Finish the from-scratch trio so nothing is a black box: BYO-9 (RAG engine), BYO-10 (ReAct agent), BYO-11 (vector DB). Ready for Phase 6B when you've shipped a deployed, evaluated RAG system and a working agent, and can explain every component.
Full guide: phase-6-llms-rag-agents.md ↗