2 min read · 381 wordsragembeddingslanggraph
Module 5: Advanced AI Engineering
Overview
Module 3 covered the RAG-relevant services in isolation; this module traces the whole pipeline end to end and fills in everything around it — embeddings, ChromaDB internals, the LangGraph concepts behind Corrective RAG, and three areas this backend deliberately (or accidentally) doesn't implement: background tasks, real conversation memory, and OpenAI-native tool calling. Every "not implemented" lesson here tells you exactly what's missing and what adding it would take.
What You'll Learn
- The six-stage RAG pipeline: upload → chunk → embed → store → retrieve → evaluate-then-generate-or-refuse
- The real embedding model and dimensionality (1536, corrected from an earlier draft's 384)
- ChromaDB's isolation model — and a real orphaned-vector resource leak found in the source
- Why Corrective RAG uses a graph specifically, not just an
ifstatement - Why
chat_with_agenthas no real conversation memory despite persisting full history
Prerequisites
- Module 4: API & Business Logic — this module goes deeper into the same endpoints and services
Lessons in This Module
| # | Lesson | What it covers |
|---|---|---|
| 14 | RAG Pipeline | The six-stage pipeline end to end |
| 15 | Embeddings | The real model, dimensionality, cost |
| 16 | Vector Databases & ChromaDB | Collections, isolation, the orphaned-vector bug |
| 17 | LangGraph & Corrective RAG | Why a graph, specifically, for this one pipeline |
| 18 | Background Tasks | Confirmed absent — what adding them would take |
| 19 | Prompt Engineering & Tool Calling | All six prompts; why this isn't OpenAI-native tool calling |
| 20 | Memory & Context Management | Why chat has no real memory |
Learning Outcomes
By the end of this module you can:
- Trace a single uploaded document through every table it touches, SQL and ChromaDB both
- Explain why ingestion and query embeddings must come from the same model
- Explain what "Corrective" adds to plain RAG and what it costs (one extra LLM call)
- Explain precisely why a chat conversation has no memory today, and what the minimal fix looks like