Crate ares_rag 

Retrieval Augmented Generation (RAG) Pipeline

This module provides the core RAG pipeline components for enhancing LLM responses with relevant context from your document collections.

Module Structure

• rag::embeddings - Dense embedding models (fastembed, 38+ models) [requires local-embeddings feature]
• rag::search - Search strategies (semantic, BM25, fuzzy, hybrid)
• rag::reranker - Cross-encoder reranking for improved relevance [requires local-embeddings feature]
• rag::chunker - Text chunking for document processing
• rag::cache - Embedding cache for avoiding recomputation

Feature Flags

The local-embeddings feature enables ONNX-based local embedding and reranking models. This feature is optional because the ONNX runtime (ort) can have build issues on some platforms, particularly Windows with certain MSVC versions.

Note: The local-embeddings feature is NOT supported on Windows MSVC due to linker errors in ort-sys. Use WSL, Linux, or macOS for local embeddings, or use remote embedding APIs.

Without local-embeddings, you can still use:

• Remote embedding APIs (OpenAI embeddings, Ollama embeddings, etc.)
• The chunker and search modules
• The cache module (if you have embeddings from elsewhere)

RAG Pipeline

The typical RAG pipeline flow:

1. Ingestion - Documents are chunked and embedded
2. Storage - Embeddings stored in vector database
3. Retrieval - Query embedded, similar chunks retrieved
4. Reranking - Cross-encoder reranks for relevance
5. Generation - LLM generates response with context

Example

use ares::rag::{embeddings::EmbeddingModel, chunker::Chunker, search::SearchStrategy};

// Embed a document
let embedder = EmbeddingModel::new("BAAI/bge-small-en-v1.5")?;
let chunker = Chunker::new(512, 50);  // chunk_size, overlap

let chunks = chunker.chunk(&document_text);
let embeddings = embedder.embed_batch(&chunks).await?;

// Search
let query_embedding = embedder.embed(&query).await?;
let results = vector_store.search("my_collection", query_embedding, 10).await?;

Embedding Models

Supports 38+ models via fastembed. Popular choices:

• BAAI/bge-small-en-v1.5 - Fast, good quality (default)
• BAAI/bge-base-en-v1.5 - Higher quality, slower
• sentence-transformers/all-MiniLM-L6-v2 - Lightweight

Re-exports

pub use config::HybridWeightsConfig;

pub use config::RagChunkingConfig;

pub use config::RagConfig;

pub use config::RagRerankingConfig;

pub use config::RagSearchConfig;

pub use config::RAGVectorConfig;

Modules

cache

Embedding Cache for RAG Pipeline

chunker

Text chunking for document processing.

config

search

Search strategies for RAG pipeline.