Crate reflex 

Reflex Cache

Episodic memory + semantic cache for LLM responses.

Reflex sits between a client (agent/server) and an LLM provider.

Request → L1 (exact) → L2 (semantic) → L3 (rerank) → Provider

Quick start

use reflex::Config;

let config = Config::from_env()?;
println!("Listening on {}", config.socket_addr());

Feature flags

Feature	Purpose
cpu	CPU-only inference (docs.rs default)
metal	Apple Silicon GPU acceleration
cuda	NVIDIA GPU acceleration
mock	Mock backends for tests/examples

Modules

• cache - Tiered cache (L1 exact + L2 semantic)
• config - Environment-backed configuration
• embedding - Embedding + reranker models
• scoring - L3 verification (cross-encoder)
• storage - Persistent cache entry storage
• vectordb - Qdrant + binary quantization utilities

Links: repo/issues at the crate repository URL.

Re-exports

pub use cache::BqSearchBackend;

pub use cache::DEFAULT_TOP_K_BQ;

pub use cache::DEFAULT_TOP_K_FINAL;

pub use cache::L2_COLLECTION_NAME;

pub use cache::L2_VECTOR_SIZE;

pub use cache::L2CacheError;

pub use cache::L2CacheResult;

pub use cache::L2Config;

pub use cache::L2LookupResult;

pub use cache::L2SemanticCache;

pub use cache::L2SemanticCacheHandle;

pub use cache::NvmeStorageLoader;

pub use cache::StorageLoader;

pub use cache::MockL2SemanticCache;mock

pub use cache::MockStorageLoader;mock

pub use cache::MockTieredCache;mock

pub use cache::L1Cache;

pub use cache::L1CacheHandle;

pub use cache::L1LookupResult;

pub use cache::REFLEX_STATUS_ERROR;

pub use cache::REFLEX_STATUS_HEADER;

pub use cache::REFLEX_STATUS_HEALTHY;

pub use cache::REFLEX_STATUS_NOT_READY;

pub use cache::REFLEX_STATUS_READY;

pub use cache::REFLEX_STATUS_STORED;

pub use cache::ReflexStatus;

pub use cache::TieredCache;

pub use cache::TieredCacheHandle;

pub use cache::TieredLookupResult;

pub use config::Config;

pub use config::ConfigError;

pub use constants::DimConfig;

pub use constants::DimValidationError;

pub use constants::validate_embedding_dim;

pub use embedding::DEFAULT_THRESHOLD;

pub use embedding::EmbeddingError;

pub use embedding::Reranker;

pub use embedding::RerankerConfig;

pub use embedding::RerankerError;

pub use embedding::SINTER_EMBEDDING_DIM;

pub use embedding::SINTER_MAX_SEQ_LEN;

pub use embedding::SinterConfig;

pub use embedding::SinterEmbedder;

pub use hashing::hash_context;

pub use hashing::hash_prompt;

pub use hashing::hash_tenant_id;

pub use hashing::hash_to_u64;

pub use lifecycle::ActivityRecorder;

pub use lifecycle::DEFAULT_IDLE_TIMEOUT_SECS;

pub use lifecycle::DEFAULT_SNAPSHOT_FILENAME;

pub use lifecycle::DehydrationResult;

pub use lifecycle::HydrationResult;

pub use lifecycle::LifecycleConfig;

pub use lifecycle::LifecycleError;

pub use lifecycle::LifecycleManager;

pub use lifecycle::LifecycleResult;

pub use lifecycle::REAPER_CHECK_INTERVAL_SECS;

pub use payload::TauqBatchEncoder;

pub use payload::TauqDecoder;

pub use payload::TauqEncoder;

pub use scoring::CrossEncoderScorer;

pub use scoring::ScoringError;

pub use scoring::VerificationResult;

pub use scoring::VerifiedCandidate;

pub use storage::CacheEntry;

pub use vectordb::bq::MockBqClient;mock

pub use vectordb::bq::BQ_BYTES_PER_VECTOR;

pub use vectordb::bq::BQ_COLLECTION_NAME;

pub use vectordb::bq::BQ_COMPRESSION_RATIO;

pub use vectordb::bq::BQ_VECTOR_SIZE;

pub use vectordb::bq::BqClient;

pub use vectordb::bq::BqConfig;

pub use vectordb::bq::DEFAULT_RESCORE_CANDIDATES;

pub use vectordb::bq::ORIGINAL_BYTES_PER_VECTOR;

pub use vectordb::bq::hamming_distance;

pub use vectordb::bq::quantize_to_binary;

pub use vectordb::MockVectorDbClient;mock

pub use vectordb::rescoring::CandidateEntry;

pub use vectordb::rescoring::DEFAULT_EMBEDDING_DIM;

pub use vectordb::rescoring::DEFAULT_TOP_K;

pub use vectordb::rescoring::EMBEDDING_BYTES;

pub use vectordb::rescoring::RescorerConfig;

pub use vectordb::rescoring::RescoringError;

pub use vectordb::rescoring::RescoringResult;

pub use vectordb::rescoring::ScoredCandidate;

pub use vectordb::rescoring::VectorRescorer;

pub use vectordb::rescoring::bytes_to_f16_slice;

pub use vectordb::rescoring::cosine_similarity_f16;

pub use vectordb::rescoring::cosine_similarity_f16_f32;

pub use vectordb::rescoring::f16_slice_to_bytes;

pub use vectordb::rescoring::f16_to_f32_vec;

pub use vectordb::rescoring::f32_to_f16_vec;

pub use vectordb::DEFAULT_COLLECTION_NAME;

pub use vectordb::DEFAULT_VECTOR_SIZE;

pub use vectordb::QdrantClient;

pub use vectordb::SearchResult;

pub use vectordb::VectorDbClient;

pub use vectordb::VectorDbError;

pub use vectordb::VectorPoint;

pub use vectordb::embedding_bytes_to_f32;

pub use vectordb::f32_to_embedding_bytes;

pub use vectordb::generate_point_id;

Modules

cache

Tiered caching infrastructure.

config

Environment-backed configuration.

constants

Cross-cutting, shared constants.

embedding

Embedding + model utilities.

hashing

Hash utilities for cache keys and identifiers.

lifecycle

Spot-instance lifecycle helpers for GCE (hydrate/dehydrate + idle reaper).

payload

Payload encoding/decoding helpers.

scoring

L3 verification via cross-encoder reranking.

storage

Storage primitives.

vectordb

Qdrant vector database integration.

Enums

TauqError

Errors returned by Tauq parsing. Top-level Tauq error type