Semantic Search
Querying S3-derived vector embeddings to find content by meaning rather than exact keyword match.
Summary
Querying S3-derived vector embeddings to find content by meaning rather than exact keyword match.
Semantic search is the retrieval layer that makes LLMs useful over S3 data. It powers the "R" in RAG — finding the most relevant S3-stored documents for a given query without requiring exact keyword matches.
- Semantic search is approximate, not exact. Results are ranked by similarity score, not matched precisely. False positives are possible and must be handled.
- Semantic search requires embedding generation as a prerequisite. You cannot search semantically without first vectorizing the S3 data.
depends_onEmbedding Model — needs vectors to searchenablesHybrid S3 + Vector Index — the retrieval mechanism for the patternaugmentsLakehouse Architecture — adds semantic retrieval to structured datascoped_toLLM-Assisted Data Systems, Vector Indexing on Object Storage
Definition
The ability to query S3-derived vector embeddings to find content by meaning rather than exact keyword match.
S3 objects cannot be searched by content natively. Semantic search, built on embeddings generated from S3 data, allows users to find relevant documents, records, or media by describing what they need in natural language.
Recent developments
- Hybrid search is the single biggest quality improvement to a naive RAG pipeline. Production 2026 systems combine semantic (vector) search with keyword (BM25) search and re-rank — pure embedding-only retrieval is now considered insufficient for high-recall use cases. Semantic-only loses to hybrid on entity-bearing queries (product codes, names, version strings). Per Redis — RAG at Scale: Production AI Systems in 2026.
- Matryoshka embeddings let truncation trade quality for cost. Matryoshka Representation Learning trains embedding models so the most semantically important information is front-loaded into the first dimensions — you can safely truncate to 256d or 512d for storage/latency wins while retaining most retrieval quality. Now standard in text-embedding-3-* and Cohere v3 families. Per Tensoria — 8 Embedding Models Compared for Production RAG 2026.
- Semantic caching cuts LLM calls 30–50%. Cache embeddings of recent queries; if a new query is >0.95 cosine-similar to a cached one, return the cached answer instead of re-running retrieval+generation. Standard pattern in production RAG stacks. Per Redis — RAG at Scale.
- text-embedding-3-small remains the default English baseline at $0.02/1M tokens. Covers 95% of general-purpose English retrieval use cases with minimal operational overhead. Switch to v3-large, Cohere embed-v3, or domain-specific models only when retrieval quality is the proven bottleneck. Per Tensoria 2026 embedding benchmark.
- Semantic chunking with 200–1,000 token windows and 10–20% overlap is the production default. Replaces fixed-size chunking; aligns chunk boundaries with paragraph/section breaks so retrieved chunks are self-contained. Per MarsDevs — RAG in AI: 2026 Production Guide.
- "RAG is not dead" is the 2026 consensus pushback. Despite agentic-retrieval hype, advanced retrieval patterns (hybrid + re-rank + late-interaction + query rewriting) still dominate in production for knowledge-base + document-search workloads. Pure agentic search wins only on code-search and structured-tool-call tasks. Per DEV — RAG Is Not Dead: Advanced Retrieval Patterns 2026.
Connections 11
Outbound 5
depends_on1enables1augments1Inbound 6
bypasses1augments1Resources 3
AWS announcement of S3 Vectors, the first cloud object store with native vector search, enabling sub-second semantic queries over billions of embeddings.
Official OpenSearch documentation for semantic search using vector embeddings, the primary open-source engine used alongside S3-backed vector stores.
AWS Big Data Blog showing how S3 Vectors integrates with OpenSearch Service for hybrid semantic + keyword search.