HAVEN searches and compares information inside encrypted databases, keeping both the data and the queries fully confidential.
HAVEN computes similarity on encrypted vectors without ever seeing the vector database and embeddings.
HAVEN searches and compares information inside encrypted databases, keeping both the data and the queries fully confidential.
HAVEN computes similarity on encrypted vectors without ever seeing the vector database and embeddings.
Lattica enables vector‑database similarity searches over encrypted embeddings and queries. We do this
without any plaintext vectors or queries touching Lattica’s servers by using FHE, which means Lattica
stores and serves only encrypted data, never plaintext vectors.
DB Administrator
End User
Vectorized DB
Query Vector
Response
Database
Encrypt
Decrypt
Deploy Encrypted DB
Encrypted Query
Encrypted Response
DB Administrator
End User
Vectorized DB
Query Vector
Response
Database
Encrypt
Decrypt
Deploy Encrypted DB
Encrypted Query
Encrypted Response
Plaintext searches assume the database operator and cloud provider are trustworthy. FHE removes that assumption. The server computes similarity scores without ever learning what’s inside, which means privacy doesn’t depend on anyone else.
Sensitive data is never exposed, dramatically simplifying regulatory risk, access controls and audit layers. An FHE-based approach inherently supports privacy regulations and data sovereignty requirements because data stays encrypted throughout processing.
Lattica’s HAVEN unlocks new use cases on data that has been off-limits due to privacy concerns. From healthcare to finance to LLMs, organizations can safely search and analyze extremely sensitive patient records, financial transactions and proprietary documents without exposing any protected information.
You generate and hold the secret key; we never do. HAVEN maintains a single‑key setup for a single organization, enabling multiple end‑users to search without managing too many keys.
For finance, healthcare, and legal data where queries and embeddings must stay private.
Face, voice, or fingerprint matching where templates stay encrypted end-to-end.
Retrieve documents for large language models while keeping embeddings encrypted
Store and compare text or image embeddings securely using encrypted vector memory.
Most “secure” approaches rely on partial encryption or trusted execution environments (TEEs) that still expose data during computation. FHE keeps both the query and database encrypted at every step. The server computes on ciphertext only, producing encrypted similarity scores that are safely decrypted by the data owner.
With you. You generate and hold the secret key.
You encrypt and decrypt on your side before and after querying, uploading the encrypted database and queries via Lattica’s client API.
FHE adds overhead, but not prohibitively. HAVEN supports representative dataset sizes and latency targets and our implementation is optimized for fast, accurate results.