Skip to main content

Pinecone Hybrid Search

Pinecone is a vector database with broad functionality.

This notebook goes over how to use a retriever that under the hood uses Pinecone and Hybrid Search.

The logic of this retriever is taken from this documentation

To use Pinecone, you must have an API key and an Environment. Here are the installation instructions.

%pip install --upgrade --quiet  pinecone-client pinecone-text
import getpass
import os

os.environ["PINECONE_API_KEY"] = getpass.getpass("Pinecone API Key:")
from langchain_community.retrievers import (
os.environ["PINECONE_ENVIRONMENT"] = getpass.getpass("Pinecone Environment:")

We want to use OpenAIEmbeddings so we have to get the OpenAI API Key.

os.environ["OPENAI_API_KEY"] = getpass.getpass("OpenAI API Key:")

Setup Pinecone

You should only have to do this part once.

Note: it's important to make sure that the "context" field that holds the document text in the metadata is not indexed. Currently you need to specify explicitly the fields you do want to index. For more information checkout Pinecone's docs.

import os

import pinecone

api_key = os.getenv("PINECONE_API_KEY") or "PINECONE_API_KEY"

index_name = "langchain-pinecone-hybrid-search"
WhoAmIResponse(username='load', user_label='label', projectname='load-test')
# create the index
dimension=1536, # dimensionality of dense model
metric="dotproduct", # sparse values supported only for dotproduct
metadata_config={"indexed": []}, # see explanation above

Now that its created, we can use it

index = pinecone.Index(index_name)

Get embeddings and sparse encoders

Embeddings are used for the dense vectors, tokenizer is used for the sparse vector

from langchain_openai import OpenAIEmbeddings

embeddings = OpenAIEmbeddings()

API Reference:

To encode the text to sparse values you can either choose SPLADE or BM25. For out of domain tasks we recommend using BM25.

For more information about the sparse encoders you can checkout pinecone-text library docs.

from pinecone_text.sparse import BM25Encoder

# or from pinecone_text.sparse import SpladeEncoder if you wish to work with SPLADE

# use default tf-idf values
bm25_encoder = BM25Encoder().default()

The above code is using default tfids values. It's highly recommended to fit the tf-idf values to your own corpus. You can do it as follow:

corpus = ["foo", "bar", "world", "hello"]

# fit tf-idf values on your corpus

# store the values to a json file

# load to your BM25Encoder object
bm25_encoder = BM25Encoder().load("bm25_values.json")

Load Retriever

We can now construct the retriever!

retriever = PineconeHybridSearchRetriever(
embeddings=embeddings, sparse_encoder=bm25_encoder, index=index

Add texts (if necessary)

We can optionally add texts to the retriever (if they aren't already in there)

retriever.add_texts(["foo", "bar", "world", "hello"])
100%|██████████| 1/1 [00:02<00:00,  2.27s/it]

Use Retriever

We can now use the retriever!

result = retriever.invoke("foo")
Document(page_content='foo', metadata={})

Help us out by providing feedback on this documentation page: