4. Advanced RAG toolkit | Notion

Retrieval Strategies

1. Hybrid search (fixing keyword failures)
2.Reranking (fixing top-k precision)
3.Query rewriting (fixing user-query weirdness)
4.HyDE (fixing query-document asymmetry)
5.Multi-query / fan-out (fixing single-shot recall)
6.Multi-hop / iterative retrieval (fixing complex questions)
7.Metadata filtering (fixing scope and ACL)
8.Parent-document / small-to-big retrieval (fixing chunk-size tension)
9.Self-querying (fixing structured-query gaps)
10.Graph RAG (preview — fixing relational questions)

The "which tool when" decision tree

This is the cheat sheet to memorize:

Symptom	Reach for
Exact-string / keyword misses	Hybrid search (BM25 + vector)
Right answer in top-50 but not top-5	Reranking
Conversational, pronoun-heavy queries	Query rewriting
Specialized domain, query-doc vocabulary gap	HyDE
Ambiguous queries needing broader recall	Multi-query fan-out
Compound / chain-of-reasoning questions	Multi-hop iterative retrieval
Mixed semantic + structured query	Self-querying
Per-user / per-tenant data isolation	Metadata filters (always, in production)
Chunk-size tension (precision vs context)	Small-to-big retrieval
Relational questions	Graph RAG

The order to apply them in a real customer engagement

You don't bolt all 10 on at once. The FDE move:

Naive RAG + metadata filters — minimum viable security and isolation
Add hybrid search — biggest quality jump for least effort
Add reranking — second biggest quality jump
Add query rewriting — fixes the conversational-UX gap