Feat/integrate vector database

[Baduc81]

• I've checked contribution guide.

Summary

Closes #586.

Adds IVectorIndex — a common interface for vector index backends — and integrates it with RetrievalResults.from_index(). This lets users build a gallery index once and query it repeatedly across batches, without reprocessing the gallery each time.

---

Changes

oml/retrieval/vector_index.py (new file)

• IVectorIndex — abstract base class with add(), search(), and __len__()
• SklearnKNNIndex — exact search via sklearn.neighbors.NearestNeighbors
• FaissIndex — exact/approximate search via faiss.IndexFlatL2 (swappable to IVF, HNSW, etc.)
• QdrantIndex — in-memory Qdrant collection via qdrant-client, no server required

oml/retrieval/retrieval_results.py

• Added RetrievalResults.from_index() classmethod
• Automatically computes gt_ids when both datasets implement ILabeledDataset

oml/retrieval/__init__.py

• Exported IVectorIndex, SklearnKNNIndex, FaissIndex, QdrantIndex

---

Usage

from oml.retrieval import RetrievalResults, FaissIndex

# Gallery is huge and fixed — process it once
dataset_gallery = ImageBaseDataset(galleries, transform=transform)
embeddings_gallery = inference(extractor, dataset_gallery, batch_size=4, num_workers=0)
index = FaissIndex(embeddings_gallery)

# Query in batches — gallery is never reprocessed
for queries in [queries1, queries2]:
    dataset_query = ImageBaseDataset(queries, transform=transform)
    embeddings_query = inference(extractor, dataset_query, batch_size=4, num_workers=0)
    rr = RetrievalResults.from_index(
        index=index,
        embeddings_query=embeddings_query,
        dataset_query=dataset_query,
        dataset_gallery=dataset_gallery,
    )
    rr.visualize_qg([0, 1], dataset_query=dataset_query, dataset_gallery=dataset_gallery, show=True)
    print(rr)

---

Notes

• search() returns L2 distances (not squared), sorted ascending — consistent with the rest of OML
• faiss, qdrant-client, and scikit-learn are treated as optional deps; ImportError is raised at runtime only if the relevant backend is instantiated without the package installed
• QdrantIndex uses an in-memory client (:memory:) so no external server is needed