Vector Stores - Semantic Stores

Vector Store

• What it stores:Vectors (also called embeddings), which are numeric representations of content such as documents, passages, chat messages, code, or images, plus metadata (ID, timestamp, source, and so on).
• What it does: Supports fast similarity search (and often hybrid search), so you can retrieve the most relevant pieces of information from the provided data.

Vector Store in Generative AI

• Powers retrieval-augmented generation (RAG): Embed knowledge sources with a vector store, retrieve the most similar chunks at query time, and then provide them to the LLM as grounded context.
• Grounds the responses: Improves relevance of retrieved information and reduces chances of using hallucinated information by grounding responses in retrieved enterprise content.

Example Use cases

• Finding relevant documents that match a user's questions.
• Powering contextual search in chatbots.

Example workflow

1. Chunk a PDF into paragraphs.
2. Create an embedding vector for each paragraph.
3. Store vectors + paragraph text + metadata in the vector store.
4. When a user asks a question, embed the question and retrieve the closest paragraphs to include as context for the model.

To use NL2SQL, you create an OCI Semantic Store resource.

A Semantic Store is backed by a vector store with structured data and includes the following DBTools connections:

• Enrichment Connection
• Query Connection

In the Console, to create a semantic store, you create a vector store with structured data. In the API, you can use the CreateSemanticStore API.

Learn about SQL Search (NL2SQL).