/seek

Search Engine Comparison Guide

Purpose: Decision framework for search engine and vector DB selection. Read when: Choosing between search engines or vector databases for a project.

Full-Text Search Engines

Selection Guide

Is your data > 100M documents?
  ├─ Yes → Elasticsearch or OpenSearch
  │   ├─ AWS-native? → OpenSearch
  │   └─ Multi-cloud / on-prem? → Elasticsearch
  └─ No
      ├─ Need Japanese support? → Elasticsearch/OpenSearch
      ├─ Developer experience priority? → Meilisearch
      ├─ Typo tolerance critical? → Typesense
      └─ Already have ES expertise? → Elasticsearch

Vector Index Strategy (Quick Reference)

Vector Databases

Notable recent changes (2024-2026):

• pgvector 0.8.x: Iterative index scans for HNSW/IVFFlat (auto-expand when needed), sparse vector (sparsevec) support, improved cost estimation — github.com/pgvector/pgvector/releases
• Pinecone serverless: Dedicated read nodes (early access), BYOC (AWS/GCP), index backup/restore GA, namespace-level metadata schema — docs.pinecone.io/release-notes/2025
• Weaviate 1.28: BlockMax WAND for faster BM25/hybrid search (experimental), RBAC (technical preview), native Japanese kagome_ja tokenizer — weaviate.io/blog/weaviate-1-28-release
• Qdrant 1.16: Tiered Multitenancy (dedicated shards for large tenants + shared shard for small), ACORN filtered vector search algorithm, Inline Storage for disk-efficient HNSW — qdrant.tech/blog/qdrant-1.16.x
• Milvus 2.5: Full-text search (BM25 on sparse vectors, no client-side preprocessing) — milvus.io/blog/introduce-milvus-2-5
• Milvus 2.6: RaBitQ 1-bit quantization (~28% memory, 4× QPS), hot-cold tiered storage (80% cost reduction), Woodpecker cloud-native WAL — milvus.io/blog/introduce-milvus-2-6

Selection Guide

Already using PostgreSQL?
  ├─ Yes, < 5M vectors → pgvector HNSW (no new infra)
  ├─ Yes, 5-50M vectors → pgvector + pgvectorscale (StreamingDiskANN)
  └─ No or > 50M vectors
      ├─ Want zero ops? → Pinecone serverless
      ├─ Heavy filtering + vector? → Qdrant (ACORN algorithm)
      ├─ Multi-modal / enterprise RBAC? → Weaviate
      ├─ Billion-scale open source? → Milvus 2.6
      ├─ Rapid prototyping? → ChromaDB
      └─ Large scale + cost-sensitive? → Milvus self-hosted or Qdrant self-hosted

Hybrid Search Capability Comparison

Key sparse/hybrid notes:

• Elasticsearch ELSER v2 (2025): available via Elastic Inference Service (EIS), first model in technical preview — elastic.co/search-labs/blog/sparse-vector-embedding
• OpenSearch neural sparse v2 (2024): distill model cuts parameters 50%, improves ingestion 1.74× on CPU — opensearch.org/blog/neural-sparse-v2-models

Cost Comparison (Approximate Monthly)

Prices are approximate and vary by region and configuration.

Migration Paths

Elasticsearch → OpenSearch

• API-compatible (mostly drop-in)
• Main differences: licensing, some plugins, version numbering
• Use snapshot/restore for data migration

Single DB → Dedicated Search

Phase 1: Add search engine alongside DB (dual-write or CDC)
Phase 2: Migrate read queries to search engine
Phase 3: Remove search-related indexes from primary DB
Phase 4: Optimize search engine configuration

pgvector → Dedicated Vector DB

Phase 1: Export embeddings to vector DB
Phase 2: Dual-read (query both, compare results)
Phase 3: Switch primary retrieval to vector DB
Phase 4: Keep pgvector as fallback / for joined queries