How to use hybrid search with Pinecone

[maxymed]

I'm having issues implementing hybrid search with LlamaIndex and Pinecone.
If I understand correctly, the sparse embedding model should be a subclass of BaseSparseEmbedding, so it can be used to generate the sparse embeddings.
So during upsert I would need to use something like :
vector_store = PineconeVectorStore( pinecone_index=pinecone_index, add_sparse_vector=True, sparse_embedding_model=sparse_embedding_model)
And during query, I don't need to explicitely generate the sparse vectors, I only use:
vector_store = PineconeVectorStore(pinecone_index=pinecone_index, text_key="_node_content", sparse_embedding_model=sparse_embedding_model) # <- points to the same vector store used for upserting vector_index = VectorStoreIndex.from_vector_store(vector_store, embed_model=Settings.embed_model) retriever = VectorIndexRetriever(index=vector_index, similarity_top_k=5, vector_store_query_mode="hybrid")
I think I'm missing something, because I don't see an alpha parameter to mix between sparse and dense, or a dense_top_k, and a sparse_top_k

[dosubot[bot]]

Hello @maxymed! I'm Dosu, and I'm here to help you solve bugs, answer questions, and guide you as you contribute. While you're waiting for a human maintainer, feel free to ask me anything!

To implement hybrid search with LlamaIndex and Pinecone, you are on the right track with using a sparse embedding model subclassed from BaseSparseEmbedding. Here are some steps and considerations to ensure correct implementation:

1. Sparse Embedding Model: Ensure your sparse embedding model is correctly implemented as a subclass of BaseSparseEmbedding. This will allow it to generate sparse embeddings during the upsert process.

2. Upsert Process: When upserting, you can indeed use the PineconeVectorStore with add_sparse_vector=True and provide your sparse_embedding_model. This setup will handle the generation and storage of sparse embeddings.

3. Query Process: During querying, you don't need to explicitly generate sparse vectors again. Your setup with PineconeVectorStore and VectorStoreIndex is correct for querying.

4. Hybrid Search Parameters: For hybrid search, you need to manage parameters like alpha, dense_top_k, and sparse_top_k. The alpha parameter is used to scale the query embedding, allowing you to mix between sparse and dense embeddings. You can apply alpha by multiplying each element of the query embedding by this value.

5. Managing Top K Results: Use dense_top_k and sparse_top_k to specify the number of top results to return for dense and sparse queries, respectively. If sparse_top_k is not set, you can use similarity_top_k as a fallback.

Here's a code snippet to illustrate these steps:

# Assuming `query` is an instance with necessary attributes
if query.query_embedding is not None:
    dimension = len(query.query_embedding)
else:
    dimension = self._pinecone_index.describe_index_stats()["dimension"]
query_embedding = [0.0] * dimension

if query.mode in (VectorStoreQueryMode.DEFAULT, VectorStoreQueryMode.HYBRID):
    query_embedding = cast(List[float], query.query_embedding)
    if query.alpha is not None:
        query_embedding = [v * query.alpha for v in query_embedding]

sparse_top_k = query.sparse_top_k or query.similarity_top_k

response = self._pinecone_index.query(
    vector=query_embedding,
    sparse_vector=pinecone_sparse_vector,  # Ensure this is set correctly
    top_k=query.similarity_top_k,  # Manage dense_top_k results
    include_values=True,
    include_metadata=True,
    namespace=self.namespace,
    filter=filter,
)

# Handle sparse results separately if needed
sparse_results = self._sparse_query_with_rank(
    query.query_str, sparse_top_k, query.filters
)

# Combine dense and sparse results
all_results = response + sparse_results

This setup ensures that hybrid search is correctly implemented by managing the query embedding, applying the alpha parameter, and setting the appropriate parameters for the Pinecone query while considering both dense and sparse results ^[1].

To continue talking to Dosu, mention @dosu.

---

^{Help Dosu learn! Give it feedback: Great Response | Irrelevant Answer | Incorrect Sources | Too Verbose | Hallucination | Bug Report | Other}