Vector Database

General definition

Vector databases are purpose-built for a problem that traditional relational and key-value databases handle poorly: finding records that are semantically close to a query even when no keyword matches exactly. They work by storing pre-computed embeddings for every piece of content in the collection, then using approximate nearest-neighbour (ANN) algorithms to return the top-N most similar vectors to a query embedding.

• Dense retrieval - matching by meaning rather than exact keywords
• Approximate nearest-neighbour (ANN) search - fast similarity lookup across millions of vectors using indexing structures such as HNSW or IVF
• Metadata filtering - combining vector similarity with traditional attribute filters (date, source, category) in a single query
• Hybrid search - blending dense vector scores with sparse keyword scores (BM25) for better precision

Popular open-source and managed options include Qdrant, Weaviate, Chroma, Milvus and pgvector (a PostgreSQL extension). Vector databases are a core component in RAG pipelines: documents are chunked, embedded, and stored at index time; at query time the query is embedded and the top matching chunks are retrieved and passed to the LLM as context.

In the Ethora ecosystem

Ethora’s AI SDK includes a RAG bot framework that connects to a vector store of your choice. You index your product documentation, knowledge base articles or support history, and the bot retrieves relevant chunks before generating a response, keeping answers grounded in your actual content rather than the model’s training data.

When deployed with the self-hosted LLM agent configuration, both the embedding model and the vector database run on your own infrastructure. That keeps proprietary documents, clinical records or financial data entirely within your environment, a common requirement for regulated buyers.