Text utilities library by Pinecone.io
Project description
Pinecone Text Client
The Pinecone Text Client is a Python package that provides text utilities designed for seamless integration with Pinecone's sparse-dense (hybrid) semantic search.
⚠️ Warning
This is a public preview ("Beta") version.
For any issues or requests, please reach out to our support team.
Installation
To install the Pinecone Text Client, use the following command:
pip install -U pinecone-text
Sparse Encoding
To convert your own text corpus to sparse vectors, you can either use BM25 or SPLADE.
BM25
To encode your documents and queries using BM25 as vector for dot product search, you can use the BM25Encoder
class.
📝 NOTE:
Our current implementation of BM25 supports only static document frequency (meaning that the document frequency values are precomputed and fixed, and do not change dynamically based on new documents added to the collection).
When conducting a search, you may come across queries that contain terms not found in the training corpus but are present in the database. To address this scenario, BM25Encoder uses a default document frequency value of 1 when encoding such terms.
Usage
For an end-to-end example, you can refer to our Quora dataset generation with BM25 notebook.
from pinecone_text.sparse import BM25Encoder
corpus = ["The quick brown fox jumps over the lazy dog",
"The lazy dog is brown",
"The fox is brown"]
# Initialize BM25 and fit the corpus
bm25 = BM25Encoder()
bm25.fit(corpus)
# Encode a new document (for upsert to Pinecone index)
doc_sparse_vector = bm25.encode_documents("The brown fox is quick")
# {"indices": [102, 18, 12, ...], "values": [0.21, 0.38, 0.15, ...]}
# Encode a query (for search in Pinecone index)
query_sparse_vector = bm25.encode_queries("Which fox is brown?")
# {"indices": [102, 16, 18, ...], "values": [0.21, 0.11, 0.15, ...]}
# store BM25 params as json
bm25.dump("bm25_params.json")
# load BM25 params from json
bm25.load("bm25_params.json")
Load Default Parameters
If you want to use the default parameters for BM25Encoder
, you can call the default
method.
The default parameters were fitted on the MS MARCO passage ranking dataset.
bm25 = BM25Encoder.default()
BM25 Parameters
The BM25Encoder
class offers configurable parameters to customize the encoding:
b
: Controls document length normalization (default: 0.75).k1
: Controls term frequency saturation (default: 1.2).- Tokenization Options: Allows customization of the tokenization process, including options for handling case, punctuation, stopwords, stemming, and language selection.
These parameters can be specified when initializing the BM25Encoder class. Please read the BM25Encoder documentation for more details.
SPLADE
Currently the SpladeEncoder
class supprts only the naver/splade-cocondenser-ensembledistil model, and follows SPLADE V2 implementation.
Usage
For an end-to-end example, you can refer to our Quora dataset generation with SPLADE notebook.
from pinecone_text.sparse import SpladeEncoder
# Initialize Splade
splade = SpladeEncoder()
# encode a batch of documents
documents = ["The quick brown fox jumps over the lazy dog",
"The lazy dog is brown",
"The fox is brown"]
document_vectors = splade.encode_documents(documents)
# [{"indices": [102, 18, 12, ...], "values": [0.21, 0.38, 0.15, ...]}, ...]
# encode a query
query = "Which fox is brown?"
query_vectors = splade.encode_queries(query)
# {"indices": [102, 18, 12, ...], "values": [0.21, 0.38, 0.15, ...]}
Dense Encoding with Sentence Transformers
For dense embedding we also provide a thin wrapper over the Sentence Transformers models hosted on hugging face. See full list of models
from pinecone_text.dense.sentence_transformer_encoder import SentenceTransformerEncoder
encoder = SentenceTransformerEncoder("sentence-transformers/all-MiniLM-L6-v2")
encoder.encode_documents(["The quick brown fox jumps over the lazy dog"])
# [[0.21, 0.38, 0.15, ...]]
encoder.encode_queries(["Who jumped over the lazy dog?"])
# [[0.11, 0.43, 0.67, ...]]
Combining Sparse and Dense Encodings for Hybrid Search
To combine sparse and dense encodings for hybrid search, you can use the hybrid_convex_scale
method on your query.
This method receives both a dense vector and a sparse vector, along with a convex scaling parameter alpha
. It returns a tuple consisting of the scaled dense and sparse vectors according to the following formula: alpha * dense_vector + (1 - alpha) * sparse_vector
.
from pinecone_text.hybrid import hybrid_convex_scale
from pinecone_text.sparse import SpladeEncoder
from pinecone_text.dense.sentence_transformer_encoder import SentenceTransformerEncoder
# Initialize Splade
splade = SpladeEncoder()
# Initialize Sentence Transformer
sentence_transformer = SentenceTransformerEncoder("sentence-transformers/all-MiniLM-L6-v2")
# encode a query
sparse_vector = splade.encode_queries("Which fox is brown?")
dense_vector = sentence_transformer.encode_queries("Which fox is brown?")
# combine sparse and dense vectors
hybrid_dense, hybrid_sparse = hybrid_convex_scale(dense_vector, sparse_vector, alpha=0.8)
# ([-0.21, 0.38, 0.15, ...], {"indices": [102, 16, 18, ...], "values": [0.21, 0.11, 0.15, ...]})
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for pinecone_text-0.4.2-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 79468c197b2fc7738c1511a6b5b8e7697fad613604ad935661a438f621ad2004 |
|
MD5 | 2f9ae13f3f821919ef5562952b937c2e |
|
BLAKE2b-256 | 886a155388699a8458db58fca434a4bf3a0622648aecb4c98785a259c0b67f49 |