Vector Databases (e.g., Pinecone, Weaviate)

Pinecone is a fully managed cloud-native vector database focused on simplicity and scale with no self-hosting option, while Weaviate is open-source and can be self-hosted or used via a managed cloud service. Weaviate also includes built-in modules for automatic vectorization of text and images, whereas Pinecone requires you to supply pre-computed embeddings.

Vector databases store high-dimensional embeddings of your documents, and when a user queries the LLM, the system retrieves the most semantically similar embeddings to ground the model's response in your proprietary data. This pattern, known as RAG, reduces hallucinations by giving the LLM relevant context it was not trained on.

Yes, Weaviate can be run locally using Docker, which is the recommended setup for development. You can also use Weaviate Embedded for quick prototyping without managing Docker containers, then migrate to Weaviate Cloud or a self-hosted production deployment when ready.

The pgvector extension lets PostgreSQL store and query vector embeddings, which is sufficient for smaller datasets and proof-of-concept RAG applications. For production workloads with millions of vectors and strict latency requirements, dedicated vector databases like Pinecone, Milvus, or Qdrant offer specialised indexing algorithms like HNSW that scale more efficiently.