Elsai Retrievers

Package: elsai-retrievers  v0.1.0

Hybrid document retrieval that combines dense (semantic) and sparse (BM25) search for more accurate results than either approach alone.

Installation

pip install --extra-index-url https://core-packages.elsai.ai/root/elsai-retrievers/ elsai-retrievers==0.1.0

Requirements: Python >= 3.9

---

HybridRetriever

Merges results from multiple retrievers into a single ranked list using Reciprocal Rank Fusion (RRF). Supports combining semantic (dense vector) retrievers with BM25 sparse search over raw document chunks.

from elsai_retrievers.hybrid_retriever import HybridRetriever
from elsai_vectordb.chromadb import ChromaVectorDb
from elsai_embeddings.azure_embeddings import AzureOpenAIEmbeddingModel

# Set up the vector store retriever (dense / semantic)
embedding_model = AzureOpenAIEmbeddingModel(...)
chroma = ChromaVectorDb(persist_directory="./db")
dense_retriever = chroma.as_retriever(
    collection_name="my_docs",
    embedding_model=embedding_model,
)

# Initialise — no constructor arguments
hybrid = HybridRetriever()

# Retrieve using semantic retrievers only
results = hybrid.hybrid_retrieve(
    retrievers=[dense_retriever],
    question="What are the key findings?",
)

for doc in results:
    print(doc.page_content)
    print(doc.metadata)

Constructor: No parameters — HybridRetriever().

hybrid_retrieve() parameters:

Parameter	Description
question	The search query string
retrievers	List of semantic retrievers — any object exposing a .retrieve() method (e.g. from as_retriever())
chunks	Optional list of raw document chunks for BM25 sparse search. Pass this to enable keyword-based matching alongside semantic search

---

Combining dense and sparse (BM25)

Pass chunks to activate BM25 alongside the semantic retrievers. Both result sets are merged with RRF before being returned.

from elsai_retrievers.hybrid_retriever import HybridRetriever

# chunks = list of documents/text chunks loaded from your corpus
hybrid = HybridRetriever()

results = hybrid.hybrid_retrieve(
    chunks=chunks,
    retrievers=[dense_retriever],
    question="What does the refund policy say?",
)

---

How hybrid retrieval works

Dense search finds semantically similar documents even when exact keywords don't match. Sparse search ensures keyword-critical documents (product names, IDs, codes) are not missed. RRF combines both ranked lists without requiring score normalisation.

---

Multiple semantic retrievers

You can pass retrievers from different vector stores — results from all of them are merged through RRF:

from elsai_retrievers.hybrid_retriever import HybridRetriever

hybrid = HybridRetriever()

results = hybrid.hybrid_retrieve(
    retrievers=[chroma_retriever, pinecone_retriever],
    question="Summarise the Q3 financial results.",
)

---

Integration with Elsai Agents

from elsai_retrievers.hybrid_retriever import HybridRetriever
from strands import tool

hybrid = HybridRetriever()

@tool
def search_docs(query: str) -> str:
    """Search the knowledge base for relevant documents."""
    docs = hybrid.hybrid_retrieve(
        retrievers=[dense_retriever],
        question=query,
    )
    return "\n\n".join(d.page_content for d in docs)

agent = Agent(tools=[search_docs])
agent("What does our refund policy say about digital purchases?")