iris-haystack¶

DocumentStore

Full Haystack 2.x protocol implementation backed by InterSystems IRIS native VECTOR(DOUBLE, N) type.

Getting Started
Semantic Search

IRISEmbeddingRetriever uses IRIS native VECTOR_COSINE with SIMD optimisation — no external vector DB required.

Embedding Retriever
Keyword Search

IRISBm25Retriever implements Okapi BM25 in-memory over the filtered document set.

BM25 Retriever
Metadata Filtering

Uses Haystack's official document_matches_filter — identical behaviour to InMemoryDocumentStore, all Python types supported.

Filtering

What is iris-haystack?¶

iris-haystack integrates InterSystems IRIS as a DocumentStore for the Haystack 2.x framework. It lets you store documents, embeddings, and metadata in IRIS and retrieve them with semantic or keyword-based search — without maintaining a separate vector database.

from haystack import Document
from intersystems_iris_haystack.document_stores import IRISDocumentStore

store = IRISDocumentStore(embedding_dim=384)

store.write_documents([
    Document(content="IRIS is a multimodel database.", meta={"category": "db"}),
    Document(content="Haystack builds LLM pipelines.",  meta={"category": "ai"}),
])

print(store.count_documents())

Why IRIS?¶

Capability	What it means for you
`VECTOR(DOUBLE, N)` column	Embeddings stored natively in SQL — no separate vector DB
`VECTOR_COSINE` function	SIMD-optimised cosine similarity computed by the database engine
ANN / HNSW index	Approximate nearest-neighbour search for large collections
SQL + JSON + globals	One platform for relational data and vectors

Stack¶

intersystems-iris-haystack
├── IRISDocumentStore          ← Haystack 2.x protocol
├── IRISEmbeddingRetriever     ← VECTOR_COSINE via SQL
└── IRISBm25Retriever          ← Okapi BM25 in-memory

Credentials are managed via Haystack Secret — never hardcoded.

Quick install¶

pip install intersystems-iris-haystack

New here?

Start with the Prerequisites page, then follow Installation and Quick Start.