Module rag 

Paper-driven two-stage RAG engine.

Pipeline (per-query):

query: &str
  → Embedder::embed()                      [plug-in]
  → [Stage 1 · MemRL]  TurboQuantIndex ANN (top_k_candidates)
  → [Stage 2 · MemRL]  exact cosine rerank (top_k_rerank)
  → [RAPO]             BFS graph expansion + vector score
  → [combined score]   vector_weight·v + graph_weight·g
  → [SuperLocalMemory] quality filter (min thresholds)
  → sort desc, take k
  → [Memex(RL)]        context BFS for each result (max_context_entities)

References:

• MemRL (2601.03192): two-stage ANN → exact rerank
• RAPO (2603.02958): graph-neighbour expansion
• Memex (2603.03561): context-entity sizing
• SuperLocalMemory (2602.13398): quality threshold filtering
• NN-RAG (2511.20333): retrieval quality over quantity

Re-exports

pub use embedder::Embedder;

pub use smart_retrieval::RetrievalWeights;

pub use smart_retrieval::SmartRetrieval;

pub use smart_retrieval::SmartSearchResult;

Modules

embedder

Embedding abstraction for the RAG pipeline.

smart_retrieval

Four-signal retrieval scoring for SmartVector.

Structs

RagConfig

Configuration knobs for the RAG pipeline. All fields have defaults tuned to the paper recommendations.

RagEngine

Hybrid RAG engine backed by SQLite.

RagResult

One result row from the RAG engine.

Enums

RagError