A vector database stores numerical embedding vectors and indexes them for fast similarity search. It lets a query find documents by meaning rather than requiring exact word matches. section 10.3.3 section 10.3.7
A semantic-search pipeline embeds documents and queries into the same vector space, compares their similarity, and ranks the closest results. Splitting documents into paragraphs or sentences improves retrieval but greatly increases the number of vectors, making brute-force comparison impractical. section 10.3.3
A production vector-search system typically combines:
- An embedding model that converts text into meaningful numerical representations. section 10.3.3
- An approximate-nearest-neighbor index that retrieves ranked matches without comparing every stored vector. section 10.3.4
- A generation or question-answering model that uses the retrieved documents as context. section 10.3.7
Choosing an implementation
Open-source options include Milvus, Weaviate, and Qdrant. PostgreSQL can add vector search through pgvector, while Elasticsearch, OpenSearch, FerretDB, and Valkey can also store vectors. Hosted services such as Pinecone, Milvus, or OpenSearch manage scaling for larger applications; a local implementation such as Knowt emphasizes private-data handling. section 10.3.4 section 10.3.7
>>> rag_pipe = Pipeline() >>> rag_pipe.add_node(component=retriever, name="Retriever", inputs=["Query"]) >>> rag_pipe.add_node(component=prompt_node, name="prompt_node", inputs=["Retriever"])
The pipeline listing shows how a retriever connects a vector-search component to a prompt node, making the database’s role in retrieval-augmented generation concrete. section 10.3.11
