Semantic search — find meaning, not just keywords.
Search across billions of documents by semantic similarity. Find "automobile" when someone searches "car," surface relevant results across language barriers, and handle natural-language queries that keyword search cannot. All at sub-10ms p50 latency from your own bucket.
Keyword search misses what users actually mean.
Synonyms
Match synonyms automatically
Semantic search finds "physician" when users type "doctor," "ML model" when they search "AI algorithm." Embeddings capture semantic relationships that keyword indexes miss.
Intent
Understand query intent
"How do I cancel my subscription?" matches documents about account termination, plan changes, and cancellation policies — not just documents containing the exact phrase.
Multilingual
Cross-language retrieval
Multilingual embedding models (like multilingual-e5) let you query in English and retrieve relevant results in French, German, or Japanese — same index, no translation layer.
Where teams deploy semantic search.
E-commerce
Product discovery at scale
Search millions of product listings by semantic similarity. Find blue denim jacket when the user searches casual summer outerwear.
Legal & compliance
Contract and case retrieval
Search clause libraries and case databases by legal concept, not literal phrase. Find relevant precedents for complex queries.
Developer tools
Code and documentation search
Search codebases and docs by functionality. Find authentication examples even when documents don't use those exact words.
Customer support
KB and ticket routing
Match incoming tickets to similar resolved issues. Surface the right knowledge base articles without keyword overlap.
Meaning-first search at any scale.
1B+
docs per namespace
<10ms
p50 latency
Any
language supported
90%+
recall@10
Common questions
How is semantic search different from keyword (full-text) search?
Full-text search matches exact terms. Semantic search matches meaning — "cheap flight" returns results about "affordable airfare" even without those exact words.
Do I need to fine-tune an embedding model?
Not usually. General-purpose models like text-embedding-3-large work well for most semantic search use cases. Fine-tune only if you need domain-specific vocabulary (medical, legal, code).
How do I handle multilingual corpora?
Use a multilingual embedding model (e.g., multilingual-e5-large). ManyVector stores whatever embeddings you provide — language handling is the model's responsibility.
What is the latency for semantic search over a billion documents?
Sub-10ms p50 on hot namespaces. Cold namespace warm-up adds 50-200ms on first access, then serves from cache.