I'm unhappy with Elasticsearch vector search performance

Sorry to hear that — Elasticsearch vector search can be great for certain workloads, but performance can get disappointing quickly depending on index size, recall settings, and how it’s configured.

If you want, I can help you diagnose it. The fastest way is to share:

1. What’s slow

   • indexing?
   • query latency?
   • throughput?
   • memory usage / GC?
   • poor recall forcing re-ranks?

2. Your setup

   • Elasticsearch version
   • vector dimensions
   • number of documents
   • shard count / replicas
   • HNSW settings (m, ef_construction, query num_candidates)
   • hardware (RAM, CPU, SSD?)

3. Your query pattern

   • exact kNN?
   • filtered kNN?
   • hybrid BM25 + vector?
   • reranking pipeline?

In the meantime, the most common causes of poor performance are:

• Too many shards for the index
• High-dimensional vectors with large corpus size
• num_candidates set too high
• Filtering after vector search instead of before / efficient filter usage
• Not enough heap / off-heap memory
• Using Elasticsearch for pure vector ANN when a dedicated vector DB would fit better

If you’d like, paste:

• your mapping,
• one sample query,
• and approximate dataset size,

and I’ll suggest concrete tuning changes.