The Hook

You’re choosing a vector DB. Pinecone, Weaviate, Qdrant, Milvus, Chroma — all claim “fast ANN search.” But under the hood, they use HNSW, DiskANN, or IVF. The index algorithm determines your RAM bill, query latency, and recall at scale. Here’s the 2024 benchmark reality.

The Thesis

Index algorithm > vector DB brand. All vector DBs wrap the same three index types: HNSW (memory-hungry, high recall), DiskANN (disk-native, 95% recall at 10% RAM), IVF (fast, recall drops at scale). Your choice should be index-first, vendor-second.

The Science

The Three Index Types

Index	Principle	RAM/1M vecs (768-dim)	Recall@10	Latency (p99)	Best For
HNSW	Hierarchical graph, greedy walk	~4-8 GB	0.95-0.99	1-5 ms	High recall, <10M vecs
DiskANN	Graph + PQ compression on SSD	~0.5-1 GB	0.90-0.97	5-20 ms	10M-1B vecs, RAM-constrained
IVF	K-means clusters + inverted lists	~1-2 GB	0.80-0.95	2-10 ms	High throughput, >100M vecs

HNSW (Hierarchical Navigable Small World)

Algorithm: Multi-layer graph. Top layer = sparse express lanes. Bottom layer = dense connections. Greedy walk from top → bottom.

Parameters:

• M (max connections): 16-64 (higher = better recall, more RAM)
• efConstruction: 100-500 (build quality)
• efSearch: 50-500 (query quality)

RAM formula: ~1.5 * M * dim * 4 bytes * num_vectors

DiskANN (Microsoft, 2021)

Innovation: Graph stored on SSD + Product Quantization (PQ) compression.

Mechanism:

1. Build HNSW in memory
2. Compress vectors with PQ (8-16 bytes/vec vs 3072 bytes)
3. Store graph on SSD, PQ codes in RAM
4. Query: load PQ codes → approximate search → re-rank with full vectors from SSD

Result: 10-20× RAM reduction, 95% recall retained.

IVF (Inverted File Index)

Algorithm: K-means clustering → assign vectors to centroids → inverted lists per centroid.

Query: Find nearest centroids → search only those lists.

Parameters:

• nlist: number of centroids (1000-100000)
• nprobe: centroids to search at query time

Trade-off: Higher nlist = smaller lists = faster search but lower recall (boundary errors).

The Benchmarks (ANN Benchmarks, 2024)

1M Vectors, 768-dim (OpenAI embeddings)

Index	Recall@10	QPS (1 thread)	RAM	Build Time
HNSW (M=32, ef=200)	0.98	1,200	4.2 GB	8 min
DiskANN (PQ=8bit)	0.96	800	0.8 GB	15 min
IVF (nlist=4096, nprobe=32)	0.92	2,500	1.5 GB	3 min

100M Vectors, 768-dim

Index	Recall@10	QPS	RAM	Notes
HNSW	0.97	200	400 GB	RAM-prohibitive
DiskANN	0.95	150	8 GB	Production choice
IVF (nlist=65536)	0.93	1,000	50 GB	High throughput

The Decision Matrix

Scenario	Choose
<1M vectors, max recall, RAM available	HNSW (Qdrant, Weaviate, Milvus)
10M-1B vectors, RAM constrained	DiskANN (Milvus, SPANN)
>100M vectors, high throughput, recall >0.9	IVF (Milvus, Vald)
Hybrid: hot vectors in RAM, cold on disk	DiskANN + HNSW tiering
Embedding dimension >1536	DiskANN (PQ scales better)

The Protocol: Choose Your Index

Step 1: Estimate Scale

vectors = docs * chunks_per_doc
dim = embedding_dim (768, 1024, 1536, 3072)
ram_gb = vectors * dim * 4 * overhead_factor

Step 2: Choose Index

Vectors	RAM Budget	Recommended
<1M	>8 GB	HNSW
1M-10M	>8 GB	HNSW
10M-100M	<32 GB	DiskANN
>100M	<100 GB	DiskANN / IVF

Step 3: Tune Parameters

HNSW:

# High recall
M = 32
ef_construction = 200
ef_search = 100

DiskANN:

# Balanced
pq_bits = 8
num_threads = 8
search_list = 100

IVF:

nlist = 4096  # sqrt(vectors) * 4
nprobe = 32   # 0.5-1% of nlist

The Counterpoint

• HNSW dominates <10M: Simpler, faster, well-tested
• DiskANN complexity: Build pipeline more complex; fewer managed services
• IVF boundary errors: Vectors near centroid boundaries lost unless nprobe high
• Filtered search: HNSW + metadata filtering = slow; IVF + filtering = faster

Practical Takeaway

Start with HNSW (Qdrant default). Monitor RAM. At 10M+ vectors or RAM pressure → migrate to DiskANN (Milvus) or IVF (Milvus/Vald). The index algorithm matters more than the vendor.

---

Sources: Malkov & Yashunin (2018) HNSW; Subramanya et al. (2021) DiskANN; Johnson et al. (2019) FAISS/IVF; ANN Benchmarks (2024); Pinecone/Weaviate/Qdrant/Milvus benchmarks (2024); FAISS documentation.