How Amplitude implemented natural language-powered analytics using Amazon OpenSearch Service as a vector database

This article details how Amplitude built Ask Amplitude, an AI-powered natural language analytics assistant, using Amazon OpenSearch Service as a vector database for semantic search and retrieval-augmented generation (RAG).

• Ask Amplitude converts natural language questions into JSON queries using LLMs and custom query engines
• Initial architecture used PostgreSQL for data and separate third-party search index for keyword search
• Evolved through four iterations: brute force cosine similarity, pgvector ANN search, dual sync to OpenSearch, and final hybrid search
• OpenSearch Service unified keyword and semantic search, eliminating need for multiple synchronization pipelines
• Final architecture reduced latency, compute requirements, and complexity by consolidating all search operations
• Extended system to index 20 million user-generated charts and dashboards for richer analytical context
• Uses HNSW product quantization and byte quantization for efficient multi-tenant vector search at scale

Amplitude's iterative journey demonstrates how combining LLMs, RAG, and unified vector search capabilities enables scalable natural language analytics while reducing architectural complexity.