RAG Search

A hybrid retrieval component for Retrieval-Augmented Generation (RAG) pipelines.

RAG Search combines lexical retrieval and semantic vector retrieval, fuses their results using Reciprocal Rank Fusion (RRF), and applies a CrossEncoder reranker to produce a final ranked set of documents.

The package works with LangChain Document objects and is designed as a modular retrieval component that can be used independently or as part of a larger RAG pipeline.

Features

• BM25 lexical retrieval
• Vector-based semantic retrieval
• Local Chroma vector store
• Configurable embedding model
• Reciprocal Rank Fusion (RRF)
• CrossEncoder reranking
• Parallel BM25 and vector retrieval
• Configurable result count
• Configurable RRF rank constant
• Configurable reranker
• Support for custom BM25 and vector retrievers

Retrieval Pipeline

The retrieval process follows this flow:

                         Query
                           │
              ┌────────────┴────────────┐
              │                         │
              ▼                         ▼
        BM25 Retrieval          Vector Retrieval
              │                         │
              └────────────┬────────────┘
                           ▼
                Reciprocal Rank Fusion
                           │
                           ▼
                    CrossEncoder
                     Reranking
                           │
                           ▼
                    Top-K Results

BM25 provides lexical matching, while vector retrieval provides semantic matching. Their results are combined using Reciprocal Rank Fusion before the final CrossEncoder reranking stage.

Installation

pip install rag-search-kit

Basic Usage

from rag_search import RAGSearch

search = RAGSearch(documents)

results = search.search(
    "What is artificial intelligence?",
    top_k=5,
)

documents should be a sequence of LangChain Document objects.

The search() method returns a list of ranked Document objects.

BM25 Retriever

The package provides a standalone BM25 retriever:

from rag_search import BM25Retriever

retriever = BM25Retriever(
    documents,
    top_k=5,
)

results = retriever.retrieve(
    "What is artificial intelligence?"
)

The BM25 retriever performs lexical matching using tokenized document text.

Vector Retriever

The package also provides a standalone vector retriever:

from rag_search import VectorRetriever

retriever = VectorRetriever(
    documents,
    top_k=5,
)

results = retriever.retrieve(
    "What is artificial intelligence?"
)

By default, the vector retriever uses:

sentence-transformers/all-MiniLM-L6-v2

and stores the document embeddings in a local Chroma vector store.

A custom LangChain-compatible embedding model can also be supplied:

retriever = VectorRetriever(
    documents,
    embedding_model=custom_embedding_model,
)

Hybrid Search

The RAGSearch class combines both retrieval approaches:

from rag_search import RAGSearch

search = RAGSearch(documents)

results = search.search(
    "What is artificial intelligence?",
    top_k=5,
)

Internally, BM25 and vector retrieval are executed concurrently. Their ranked results are then combined using Reciprocal Rank Fusion (RRF), followed by CrossEncoder reranking.

Configuration

Result Count

The default retrieval limit is 5.

search = RAGSearch(documents)

results = search.search(
    "What is artificial intelligence?",
    top_k=10,
)

RRF Rank Constant

The default RRF rank constant is 60.

search = RAGSearch(
    documents,
    rank_constant=60,
)

Reranker

The default CrossEncoder reranker is:

BAAI/bge-reranker-v2-m3

A different CrossEncoder model can be supplied:

search = RAGSearch(
    documents,
    reranker="your-reranker-model",
)

Custom Retrievers

RAGSearch also allows custom BM25 and vector retriever instances:

search = RAGSearch(
    documents,
    bm25_retriever=custom_bm25,
    vector_retriever=custom_vector,
)

This allows the hybrid search layer to be composed with customized retrieval implementations.

Components

The package exposes three main components:

Component	Purpose
BM25Retriever	Lexical retrieval
VectorRetriever	Semantic vector retrieval
RAGSearch	Hybrid retrieval, RRF fusion, and reranking

Dependencies

The retrieval implementation uses components from the LangChain ecosystem together with:

• Chroma
• Hugging Face embeddings
• BM25
• Ranx for Reciprocal Rank Fusion
• Sentence Transformers for CrossEncoder reranking

Package dependencies are defined in pyproject.toml.

Role in a RAG Pipeline

RAG Search is intended to operate after document ingestion and chunking:

Documents
    ↓
Document Ingestion
    ↓
Cleaning & Chunking
    ↓
RAG Search
    ↓
Retrieved Context
    ↓
LLM

It focuses specifically on retrieval and ranking. Document loading and chunking are handled separately.

Status

This is the initial version of the RAG Search component, developed as a modular retrieval building block for a larger RAG system.

License

This project is licensed under the MIT License. See the LICENSE file for details.