
Most common position in a pipeline	Before a `DocumentWriter` in an indexing pipeline
Mandatory run variables	“documents”: A list of documents
Output variables	“documents”: A list of documents (enriched with sparse embeddings)
API reference	FastEmbed
GitHub link	https://github.com/deepset-ai/haystack-core-integrations/tree/main/integrations/fastembed

To compute a sparse embedding for a string, use the FastembedSparseTextEmbedder.

Overview

FastembedSparseDocumentEmbedder computes the sparse embeddings of a list of documents and stores the obtained vectors in the sparse_embedding field of each document. It uses sparse embedding models supported by FastEmbed.

The vectors calculated by this component are necessary for performing sparse embedding retrieval on a set of documents. During retrieval, the sparse vector representing the query is compared to those of the documents to identify the most similar or relevant ones.

Compatible models

You can find the supported models in the FastEmbed documentation.

Currently, supported models are based on SPLADE, a technique for producing sparse representations for text, where each non-zero value in the embedding is the importance weight of a term in the BERT WordPiece vocabulary. For more information, see our docs that explain sparse embedding-based Retrievers further.

Installation

To start using this integration with Haystack, install the package with:

pip install fastembed-haystack

Parameters

You can set the path where the model will be stored in a cache directory. Also, you can set the number of threads a single onnxruntime session can use:

cache_dir= "/your_cacheDirectory"
embedder = FastembedSparseDocumentEmbedder(
	model="prithvida/Splade_PP_en_v1",
	cache_dir=cache_dir,
	threads=2
)

If you want to use the data parallel encoding, you can set the parameters parallel and batch_size.

If parallel > 1, data-parallel encoding will be used. This is recommended for offline encoding of large datasets.
If parallel is 0, use all available cores.
If None, don't use data-parallel processing; use default onnxruntime threading instead.

👍
If you create both a Sparse Text Embedder and a Sparse Document Embedder based on the same model, Haystack utilizes a shared resource behind the scenes to conserve resources.

Embedding Metadata

Text documents often include metadata. If the metadata is distinctive and semantically meaningful, you can embed it along with the document's text to improve retrieval.

You can do this easily by using the sparse Document Embedder:

from haystack.preview import Document
from haystack_integrations.components.embedders.fastembed import FastembedSparseDocumentEmbedder

doc = Document(text="some text",
							 metadata={"title": "relevant title",
												 "page number": 18})

embedder = FastembedSparseDocumentEmbedder(
	model="prithvida/Splade_PP_en_v1",
	metadata_fields_to_embed=["title"]
)

docs_w_sparse_embeddings = embedder.run(documents=[doc])["documents"]

Usage

On its own

from haystack.dataclasses import Document
from haystack_integrations.components.embedders.fastembed import FastembedSparseDocumentEmbedder
document_list = [
	Document(content="I love pizza!"), 
	Document(content="I like spaghetti")
]

doc_embedder = FastembedSparseDocumentEmbedder()
doc_embedder.warm_up()

result = doc_embedder.run(document_list)
print(result['documents'][0])

# Document(id=..., 
# 				 content: 'I love pizza!', 
#					 sparse_embedding: vector with 24 non-zero elements)

In a pipeline

Currently, sparse embedding retrieval is only supported by QdrantDocumentStore.
First, install the package with:

pip install qdrant-haystack

Then, try out this pipeline:

from haystack import Document, Pipeline
from haystack.components.writers import DocumentWriter
from haystack_integrations.components.retrievers.qdrant import QdrantSparseEmbeddingRetriever
from haystack_integrations.document_stores.qdrant import QdrantDocumentStore
from haystack.document_stores.types import DuplicatePolicy
from haystack_integrations.components.embedders.fastembed import FastembedDocumentEmbedder, FastembedTextEmbedder

document_store = QdrantDocumentStore(
    ":memory:",
    recreate_index=True,
    use_sparse_embeddings=True
)

documents = [
    Document(content="My name is Wolfgang and I live in Berlin"),
    Document(content="I saw a black horse running"),
    Document(content="Germany has many big cities"),
    Document(content="fastembed is supported by and maintained by Qdrant."),
]

sparse_document_embedder = FastembedSparseDocumentEmbedder()
writer = DocumentWriter(document_store=document_store, policy=DuplicatePolicy.OVERWRITE)

indexing_pipeline = Pipeline()
indexing_pipeline.add_component("sparse_document_embedder", sparse_document_embedder)
indexing_pipeline.add_component("writer", writer)
indexing_pipeline.connect("sparse_document_embedder", "writer")

indexing_pipeline.run({"sparse_document_embedder": {"documents": documents}})

query_pipeline = Pipeline()
query_pipeline.add_component("sparse_text_embedder", FastembedSparseTextEmbedder())
query_pipeline.add_component("sparse_retriever", QdrantSparseEmbeddingRetriever(document_store=document_store))
query_pipeline.connect("sparse_text_embedder.sparse_embedding", "sparse_retriever.query_sparse_embedding")

query = "Who supports fastembed?"

result = query_pipeline.run({"sparse_text_embedder": {"text": query}})

print(result["sparse_retriever"]["documents"][0])  # noqa: T201

# Document(id=...,
#  content: 'fastembed is supported by and maintained by Qdrant.',
#  score: 0.758..)

Additional References

🧑‍🍳 Cookbook: Sparse Embedding Retrieval with Qdrant and FastEmbed