Awesome Search KG

OpenSearch

5 min read

OpenSearch

Open-source distributed search and analytics engine forked from Elasticsearch 7.10.2 in 2021 by AWS, after Elastic changed its license from Apache 2.0 to SSPL/Elastic License. Licensed under Apache 2.0 and governed by a community that includes the Linux Foundation. Primary managed offering: Amazon OpenSearch Service.

What It Does

OpenSearch indexes documents as JSON and exposes a REST API for full-text search, aggregations, filtering, and vector ANN search. Architecture mirrors Elasticsearch at the shard/node level — same Lucene foundation, same distributed replication model. Horizontally scalable.

Key capabilities:

  • Full-text search — BM25-based relevance scoring via Lucene
  • Dense vector search — k-NN plugin with HNSW; supports Faiss, nmslib, and native Lucene backends; up to 16,000 dimensions
  • Hybrid search — BM25 + kNN with normalization pipeline; native RRF support
  • Neural sparse retrieval — learned sparse encoding (open-source counterpart to ELSER)
  • Late interaction reranking — native lateInteractionScore (3.3+) for ColBERT/ColPali multi-vector rerank; see Late Interaction in OpenSearch
  • Relevance Agent — conversational, multi-agent relevance diagnosis and query-DSL tuning in Dashboards, experimental in 3.6; see OpenSearch Relevance Agent
  • Agentic memory — memory containers bundling four memory types (sessions, working, long-term, history) with LLM extraction strategies and namespaces, under /_plugins/_ml/memory_containers/; introduced in 3.3, and built for external agent frameworks (LangChain, LangGraph) as well as OpenSearch’s own. Context management for long histories is dated to 3.5 in Bouafif’s talk. See Agentic Memory
  • Search pipelines — composable request/response processors for ML inference and score normalization
  • Aggregations — significant_terms, facets, histograms; foundational for Wormhole Vectors traversal
  • PPL (Piped Processing Language) — SQL-like analytics query language

Retrieval Options

ModeQuery typeNotes
BM25match, multi_matchDefault; no ML dependency
Dense semanticknnk-NN plugin; choice of Faiss / nmslib / Lucene engine
Hybridhybrid + normalization pipelineBuilt-in RRF and min-max normalization
Neural sparseneural_sparseRequires a sparse encoding model

The hybrid query type matters more than it looks: it normalizes each branch into a comparable range before combining, which a hand-rolled bool query with two should clauses does not — that construct sums raw scores from incomparable scales. The intuitive path adds, the purpose-built path normalizes, and most teams don’t discover the difference until retrieval quality breaks. See Hybrid Fusion Failure - BM25 Displacing Reference Documents.

Vector Search: k-NN Plugin Backends

EngineNotes
FaissFacebook’s library; strong at scale; GPU support
nmslibHNSW; good recall/speed balance; original plugin default
LuceneNative HNSW; no plugin overhead; best for smaller indexes

AWS Integration

Deployed as Amazon OpenSearch Service, OpenSearch gains:

  • Native IAM, VPC, CloudWatch, and Kinesis integration
  • Direct connector to Amazon Bedrock for embedding inference and RAG generation
  • Managed scaling, snapshots, and patching

See Innovating Search Experience with Amazon OpenSearch and Amazon Bedrock.

Advanced Retrieval: Wormhole Vectors

Wormhole Vectors exploit OpenSearch’s k-NN plugin and significant_terms aggregations together to traverse between dense, sparse, and behavioral vector spaces — beyond what standard RRF hybrid search achieves.

# Dense query → aggregate statistically significant keywords from results
response = client.search(
    index="products",
    body={
        "size": 50,
        "query": {"knn": {"embedding_vector": query_vector, "k": 50}},
        "aggs": {
            "wormhole_keywords": {
                "significant_terms": {"field": "description_keywords", "size": 10}
            }
        }
    }
)

See Wormhole Vectors Beyond Hybrid Search in OpenSearch — Dima Kan (Aiven).

Performance vs. Elasticsearch

2024–2025 benchmarks show Elasticsearch consistently faster:

  • Text search / aggregations: 40–140% faster
  • Vector search: 2–12x faster

The gap is largest on complex, high-throughput workloads. At smaller scales the difference narrows and OpenSearch is often sufficient. See Elasticsearch vs OpenSearch for the full comparison.

Notable Implementations in This Vault

  • Dima Kan (Aiven) — Wormhole Vectors in production; SKG traversal across sparse, dense, and behavioral spaces
  • Canva — one of three parallel candidate generators alongside Solr and SageMaker (see Canva Search Pipeline Part II)
  • Elasticsearch — upstream project; faster but proprietary-licensed above 7.10
  • Metarank — external LambdaMART secondary re-ranker that integrates with OpenSearch (Remote Ranker Plugin RFC)
  • Search Relevance Workbench — native (3.1+) offline relevance-evaluation tool: query sets, judgments, experiments, and hybrid-search tuning inside OpenSearch Dashboards; see Relevance Evaluation Tools Compared
  • OpenSearch Dashboards — visualization layer (Kibana fork)

Articles

Comparison

Graph View

Backlinks