Index Tuning
pgvector IVFFlat Lists Tuning Guide: Probes and Recall
Set pgvector IVFFlat lists to rows/1000 up to 1M rows and sqrt(rows) above, probes to sqrt(lists), then measu…
VectorOpsReport — Vector Database HNSW Index & ANN Performance Report
Vector Database HNSW Index & ANN Performance Report
VectorOpsReport
Independent · 2026
Independent · 2026
Scaling
How to Shard a Large Vector Index Without Losing Recall
Hash-partition by ID, size shards to fit RAM, fan out with full top-k per shard and merge. Shard counts, routing and fan-out tail latency math.
Read the article →Latest articles
Guide
What Is Reranking in Vector Search? A Practical Guide
Learn how reranking improves vector search relevance, measure nDCG and candidate recall, and instrument a PyT…
Vector Search
Product Quantization Explained for Vector Search
Learn how product quantization compresses embeddings, how IVF-PQ changes recall, and how to evaluate a Faiss …
Indexing
HNSW M Parameter Tuning: Recall, Memory, and Latency
The guide explains how M affects graph connectivity and memory, maps engine-specific settings, and shows a sw…
Guide
How to Choose a Vector Database for Production
The guide compares pgvector, Qdrant, Milvus, Weaviate, Pinecone, and Elasticsearch by recall, filters, latenc…
Fundamentals
Cosine Similarity vs Dot Product: How Rankings Differ
Vector norms determine when cosine similarity and dot product rank results identically and when magnitude cha…
The field guide
Vector search breaks in a handful of predictable ways: an index that will not fit in memory, recall that quietly drops, a distance metric that does not match the embedding model, or a database chosen before the constraints were known. VectorOpsReport works those problems from the engine documentation up, with the arithmetic to size an index and the harness to measure it. New to the topic? Start with the vector search fundamentals, then size a build with the HNSW index RAM calculator.