DuckDB Document Store for Haystack

[!NOTE] This project is a proof of concept - use at your own risk. The code may be susceptible to bugs and security issues (such as SQL injection), proceed with caution.

A DuckDB-backed document store for Haystack with HNSW vector search via DuckDB's VSS extension. It supports:

• Dense embedding storage with HNSW indexing (cosine similarity, Euclidean distance, or inner product distance)
• Filtering with Haystack-style filter dictionaries
• In-memory operation or persistence via a DuckDB database file on disk

Installation (GitHub)

Use uv to install directly from the repository:

uv pip install "duckdb-haystack @ git+https://github.com/AdrianoKF/duckdb-haystack.git"

Usage

1) DocumentStore CRUD example

from haystack import Document

from haystack_integrations.document_stores.duckdb import DuckDBDocumentStore, document_store

store = DuckDBDocumentStore(
    database=":memory:",
    embedding_dim=3,
    similarity_metric="cosine",
)

store.write_documents(
    [
        Document(id="doc-1", content="DuckDB is fast.", embedding=[0.1, 0.0, 0.9], meta={"source": "notes"}),
        Document(id="doc-2", content="Haystack pipelines are modular.", embedding=[0.2, 0.1, 0.8]),
    ]
)

print("Total document count:", store.count_documents())

filters = {"field": "meta.source", "operator": "==", "value": "notes"}
filtered = store.filter_documents(filters=filters)
print("Filtered documents:", [doc.id for doc in filtered])

store.delete_documents(document_ids=["doc-2"])
print("After deletion:", store.filter_documents())

2) Retrieval with DuckDBRetriever in a pipeline

from haystack import Document, Pipeline
from haystack.components.embedders import SentenceTransformersDocumentEmbedder, SentenceTransformersTextEmbedder

from haystack_integrations.document_stores.duckdb import DuckDBDocumentStore
from haystack_integrations.retrievers.duckdb import DuckDBRetriever

store = DuckDBDocumentStore(database=":memory:", embedding_dim=384)

doc_embedder = SentenceTransformersDocumentEmbedder(model="sentence-transformers/all-MiniLM-L6-v2")
doc_embedder.warm_up()
documents = [
    Document(content="DuckDB stores vectors in Float arrays backed by an HNSW index."),
    Document(content="DuckDB is an analytical in-process SQL database management system."),
    Document(content="Haystack offers composable pipelines."),
]
documents = doc_embedder.run(documents=documents)["documents"]
store.write_documents(documents)

query_embedder = SentenceTransformersTextEmbedder(model="sentence-transformers/all-MiniLM-L6-v2")
retriever = DuckDBRetriever(document_store=store)

pipeline = Pipeline()
pipeline.add_component("query_embedder", query_embedder)
pipeline.add_component("retriever", retriever)
pipeline.connect("query_embedder.embedding", "retriever.query_embedding")

result = pipeline.run(data={"query_embedder": {"text": "How does DuckDB store vectors?"}})

print(result["retriever"]["documents"][0].content)

License

duckdb-haystack is distributed under the terms of the Apache-2.0 license.