Retrieval-Augmented Generation (RAG)

3 steps:

1. Indexing
2. Retrieval
3. Generation

Some common techniques / terminology that you should know: Query Construction (how do you augment our initial query to get better search)

• Multi-query: where you rewrite the query N-ways through LLM, then redo retrieval N-times
• RAG-fusion: builds on multi-query by adding sophisticated ranking step by adding RRF
• Query Translation
  • Sub-question
  • Step-back
  • Using LLM to generate domain-specific language (usually in the form of JSON, similar to routing?)
• HyDE
• Routing
  • I was introduced to structured outputs that you can specify
  1. Logical routing
  2. Semantic routing

Indexing (this is where I don’t fully understand?)

• Multi-representation
• RAPTOR
• ColBERT (you can use ratatouille to do this)
• Active RAG