Vector Database Indexing: HNSW vs IVF

Not all vector indexes are created equal. Here’s my comparison after running both in production.

HNSW (Hierarchical Navigable Small World)

import hnswlib

index = hnswlib.Index(space='cosine', dim=1536)
index.init_index(max_elements=1000000, ef_construction=200, M=16)

index.add_items(vectors, ids)

labels, distances = index.knn_query(query_vector, k=10)

Pros: Fast queries, no training needed
Cons: High memory usage, slow inserts

IVF (Inverted File Index)

import faiss

quantizer = faiss.IndexFlatL2(1536)
index = faiss.IndexIVFFlat(quantizer, 1536, nlist=1000)
index.train(training_vectors)

index.add(vectors)

index.nprobe = 10
distances, labels = index.search(query_vector, k=10)

Pros: Lower memory, fast bulk inserts
Cons: Requires training, slightly lower recall

Benchmark Results (1M vectors, 1536 dims)

Metric	HNSW	IVF
Query Latency	2ms	5ms
Memory	6GB	2GB
Insert Time	45min	8min
Recall@10	98%	94%

My recommendation: HNSW for < 10M vectors, IVF-PQ for larger scales.