The scenario
Atlas RAG is a reference implementation of modern retrieval: hybrid search with reranking, hypothetical document embeddings, corrective and self-reflective RAG loops, Text-to-SQL for structured questions, semantic caching and guardrails — all orchestrated in LangGraph. The point is not to use everything, but to measure each technique against a baseline and keep only what earns its latency.
System architecture
01 Question Docs or data?
02 Router RAG or Text-to-SQL
03 Hybrid retrieval BM25 + vectors + rerank
04 Self-correction Grade → retry
05 Answer Cached + guarded
High-level flow of Atlas RAG, from input to output.
What you will learn
- Hybrid retrieval and reranking
- HyDE, corrective RAG and self-RAG loops
- Routing between documents and SQL
- Caching and guardrails
- Ablation-style evaluation of each technique
What you will build
- Configurable retrieval graph
- Text-to-SQL branch with safety checks
- Semantic cache
- Evaluation dashboard comparing techniques
Build plan
- Baseline RAG + eval set
- Hybrid + rerank
- HyDE and corrective loops
- Text-to-SQL branch
- Cache, guardrails, report