Six stages, and any one of them can lie to you.

The whole course in one picture. Every later lesson zooms into one box of this pipeline.

Spaced recall (from your Knowledge Bases course, no peeking): (1) which search style nails an exact ID like a policy number? (2) what turns text into numbers you can compare by meaning? Answers: keyword/BM25 · embeddings. Both come back today. Refreshers: KB Lesson 2, glossary.

The one idea: RAG (Retrieval-Augmented Generation) is a librarian pipeline: documents go in one end, cited answers come out the other. Answer quality is capped by the worst stage, not the best model. Enterprise buyers pay for the pipeline being trustworthy, not for the LLM being clever.

The pipeline

1 Loadread the raw documents (PDFs, pages) into text

2 Chunksplit into paragraph-sized pieces

3 Embedturn each chunk into meaning-numbers

4 Storesave vectors in a vector database (Chroma)

5 Retrieveembed the question, fetch the nearest chunks

6 GenerateLLM answers from those chunks, with citations

Stages 1-4 run once per document set (ingest). Stages 5-6 run on every customer question. That split matters: ingest mistakes are invisible until a customer hits them weeks later.

What breaks at each stage (the Samsung Life edition)

This table is the sales asset. A prospect trusts you when you can name the failure modes before they happen.

Stage	How it lies	Insurance example
Load	PDF tables and layouts get mangled into word soup	A premium-by-age table reads as one garbled line; the agent quotes the wrong premium
Chunk	A clause is cut in half at a chunk boundary	"Flood damage is covered" and "except in basement units" land in different chunks
Embed	Wrong embedding model for the language	Samsung Life policies are Korean; an English-only embedder returns junk. Same lesson as Thai at DuckyCutz
Store	The snapshot goes stale	A policy revision ships; the store still serves last quarter's terms
Retrieve	Meaning-search misses exact identifiers	"Policy S-240-113" needs keyword match, not similarity. Hybrid search fixes this (you learned it in KB Lesson 2)
Generate	The LLM cites a clause that is not in the retrieved chunks	Compliance nightmare. Fix: force citations to retrieved chunk IDs, validate in code

Why DuckyCutz has none of this, on purpose

The barbershop agent answers questions about live state ("is 3pm free with Ohm?"). A vector store is a snapshot, so stage 4 would lie within minutes and cause double-bookings. Insurance answers live in documents that change quarterly, not by the minute. Snapshot data is exactly what RAG is for. One sentence to keep: RAG for documents, live API calls for live state, and a real agent for an enterprise usually needs both as separate tools.

Where LangGraph fits

Nothing above is LangGraph. LangGraph is the flowchart around the pipeline: identify the customer, fetch their policy, run retrieval, draft the answer, and pause for a human before anything touching money goes out. That human-approval gate is the enterprise feature. It arrives in Lesson 4, once retrieval works.

The road to the demo

Pipeline anatomy (today, done)
Build the corpus + chunk it: write the Samsung Life sample policy set, split it in Python (~30 min)
Embed, store, retrieve: Chroma + a retrieval quality check you can show a prospect (~40 min)
LangGraph: state, nodes, edges, the human-approval interrupt (~40 min)
Capstone: full demo agent, question in, cited answer out + the pitch script (~1 hr)

Prove it stuck

A coverage clause exists in the PDF but retrieval never returns it intact. Which stage most likely failed?

The Samsung Life corpus is Korean and retrieval returns junk for every query. First thing to swap?

Why is RAG the wrong tool for "is 3pm free with Ohm?"

Read this next (primary source)

★ LangChain's official RAG tutorial, Part 1. It walks the exact six stages above with runnable Python. ~20 min read; we build our own version of it in Lessons 2-3, so skimming it now pays off twice.

Ask your teacher. Try: "show me a real chunk boundary failure on an insurance clause", "which multilingual embedder handles Korean well?", or "what would the human-approval gate look like in the demo?" Anything unclear, ask before Lesson 2.

Course start Next → Lesson 2: Build the corpus + chunk it