Module store 

Vector store abstraction for chunk persistence and hybrid search.

The VectorStore trait decouples the RAG pipeline from any specific storage backend. Two implementations are provided behind feature flags:

• internal (default) — pure Rust, zero native deps, MessagePack on disk
• lancedb — LanceDB-backed hybrid BM25 + vector search

Structs

BruteForceStore

Pure‑Rust brute‑force vector store backed by MessagePack on disk. Enabled by the internal feature (on by default).

HybridRrfRanker

Cosine-similarity + BM25 fused via Reciprocal Rank Fusion (k = 60).

LlmReranker

LLM-based re-ranker: delegates candidate retrieval to an inner ranker, then asks an LLM to re-order the candidates by relevance to the query.

MmrDiversityRanker

Maximal Marginal Relevance (MMR) diversity re-ranker.

ScoredChunk

Result of a hybrid search.

StoredChunk

A single chunk with its embedding, ready to be stored.

WeightedFusionRanker

Weighted linear fusion of cosine similarity and BM25.

Traits

Ranker

Pluggable chunk ranking strategy.

VectorStore

Backend-agnostic chunk storage with hybrid BM25 + vector search.

Functions

embed_and_insert

Helper: convert embedded (text, Vec<f32>) pairs into StoredChunks keyed by source file, then insert into the store.

open_store

Open or create a vector store in the given folder. Uses the built‑in brute‑force store (MessagePack on disk).