Struct DualRetriever 

pub struct DualRetriever<L: LLMProvider, E: EmbeddingProvider, V: VectorBackend, S: StorageBackend> { /* private fields */ }

Dual retriever: fast search + LLM reasoning.

TigerStyle: Generic over LLM and storage for sim/production flexibility.

Example

use umi_memory::retrieval::{DualRetriever, SearchOptions};
use umi_memory::llm::SimLLMProvider;
use umi_memory::embedding::SimEmbeddingProvider;
use umi_memory::storage::{SimStorageBackend, SimVectorBackend};
use umi_memory::dst::SimConfig;

#[tokio::main]
async fn main() {
    let llm = SimLLMProvider::with_seed(42);
    let embedder = SimEmbeddingProvider::with_seed(42);
    let vector = SimVectorBackend::new(42);
    let storage = SimStorageBackend::new(SimConfig::with_seed(42));
    let retriever = DualRetriever::new(llm, embedder, vector, storage);

    // Deep search with query rewriting + vector search
    let result = retriever
        .search("Who works at Acme?", SearchOptions::default())
        .await
        .unwrap();
}

Implementations

impl<L: LLMProvider, E: EmbeddingProvider, V: VectorBackend, S: StorageBackend> DualRetriever<L, E, V, S>

pub fn new(llm: L, embedder: E, vector: V, storage: S) -> Self

Create a new dual retriever.

pub async fn search( &self, query: &str, options: SearchOptions, ) -> Result<SearchResult, RetrievalError>

Search with dual retrieval strategy.

Arguments

• query - Search query
• options - Search options (limit, deep_search, time_range)

Returns

SearchResult with entities, query info, and metadata.

Errors

Returns RetrievalError if query is empty, too long, or limit is invalid.

Graceful Degradation

If LLM fails during query rewriting, falls back to fast search only.

pub async fn rewrite_query(&self, query: &str) -> Vec<String>

Rewrite query into search variations using LLM.

Arguments

• query - Original search query

Returns

Vector of query variations (always includes original).

Graceful Degradation

Returns only the original query if LLM fails.

pub fn merge_rrf(&self, result_lists: &[&Vec<Entity>]) -> Vec<Entity>

Merge results using Reciprocal Rank Fusion.

RRF score: sum(1 / (k + rank)) for each list the document appears in. Documents appearing in multiple lists get higher scores.

Arguments

• result_lists - Slice of entity vectors to merge

Returns

Merged and deduplicated entities, sorted by RRF score.

pub fn llm(&self) -> &L

Get a reference to the underlying LLM provider.

pub fn storage(&self) -> &S

Get a reference to the underlying storage backend.

Trait Implementations

impl<L: Debug + LLMProvider, E: Debug + EmbeddingProvider, V: Debug + VectorBackend, S: Debug + StorageBackend> Debug for DualRetriever<L, E, V, S>

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.