Region set (BED) search combining both metadata and genomic regions

Metadata embedding vector backend

This vector beckend stored the embedding vectors of region set metadata annotations, which are encoded by open-source text model (SentenceTransformers, etc.). The payload of each metadata annotation vector must contain the storage ids of region set that the annotation matches.

Example code

from geniml.search.backends import BiVectorBackend, QdrantBackend
from geniml.search.interfaces import BiVectorSearchInterface

# 2 required backends
text_backend = QdrantBackend(dim=384)
bed_backend = QdrantBackend()

# load vectors and payloads
bed_backend.load(vectors=np.array(bed_vecs), payloads=bed_payloads)
text_backend.load(vectors=np.array(text_embeddings), payloads=text_payloads)

# the search backend
search_backend = BiVectorBackend(text_backend, bed_backend)


# the search interface
search_interface = BiVectorSearchInterface(
    backend=search_backend, query2vec="sentence-transformers/all-MiniLM-L6-v2"
)


# actual search
result = search_interface.query_search(
    query="lung cancer cell lines",
    limit=10,
    with_payload=True,
    with_vectors=False,
    p=1.0,
    q=1.0,
    distance = False # QdrantBackend returns similarity as the score, not distance
)