qdrant-search-speed-optimization

First check whether a repeated query runs faster, which indicates memory pressure, and test with `with_payload: false` to isolate payload retrieval costs. Common fixes include tuning HNSW parameters, enabling in-memory scalar quantization, and reducing vector dimensionality with Matryoshka models.

Reduce `default_segment_number` to 2, use the batch search API instead of single queries, enable quantization to cut CPU cost, and add replicas to distribute read load. Fewer segments improve throughput rather than more.

Filtered search is slow when the filtered field lacks a payload index, forcing Qdrant to scan raw payloads. Create a payload index, use `is_tenant=true` for primary filters, try the ACORN algorithm for complex filters, and avoid nested conditions as primary filters.

The optimizer is likely still indexing new segments. Test with `indexed_only=true`; if much faster, indexing is incomplete. Reduce `optimizer_cpu_budget` to reserve CPU for queries and set `prevent_unoptimized=true` to defer unindexed data.

Avoid keeping quantized vectors out of RAM, putting HNSW on disk for latency-sensitive workloads, increasing segment count for throughput, and indexing every payload field. Also check optimizer status before blaming Qdrant itself.