3.2 Document Vectorization Concepts supporting RAG

Explains document vectorization concepts that support RAG.

Topics:

• Vector
  
• Vector Database
  
• Vector Distance
  
• Vector Index
  
• Vector Store
  
• Embedding Model
  
• Transformer
  
• ONNX Format
  
• ONNX Runtime
  
• Semantic Similarity Search
  
• Similarity Threshold
  

3.2.1 Vector

In the context of semantic similarity search, a vector is a mathematical representation that captures the semantic meaning of data points, such as words, documents, or images, in a multi-dimensional space.

In the context of Select AI, vectors support retrieval augmented generation by capturing the meaning of text content to enable fast semantic retrieval from the database.

3.2.2 Vector Database

A database that stores vector embeddings, which are mathematical representations of data points used in AI applications to support efficient semantic similarity search. Oracle Autonomous AI Database and Oracle AI Database serve as a vector database with optimized vector indexes.

In Select AI, the vector database component (powered by Oracle AI Vector Search) indexes embeddings generated from enterprise data. This enables natural language queries to retrieve semantically similar results, improves relevance for AI-powered search and RAG workflows, and provides seamless integration with Oracle Cloud environments.

3.2.3 Vector Distance

Vector distance measure the similarity or dissimilarity between feature vectors by calculating the distance between them in a multidimensional space.

3.2.4 Vector Index

A vector index organizes and stores vectors to enable efficient similarity search and retrieval of related data.

3.2.5 Vector Store

A vector store includes systems that store, manage, and enable semantic similarity search involving vector embeddings. This includes standalone vector databases and Oracle AI Database 26ai AI Vector Search.

3.2.6 Embedding Model

An AI model that converts input data into vector embeddings to capture semantic relationships, often used in tasks like language understanding and image recognition. Select AI uses embedding models to compute embeddings for your documents, tables, and query text. These embeddings power semantic search, RAG workflows, similarity scoring, and relevance ranking inside Autonomous AI Database.

3.2.7 Transformer

A type of deep learning model architecture commonly used for natural language processing tasks, such as vector embedding generation or text generation and translation. In Select AI, transformer-based LLMs drive the conversion of user queries into SQL queries that can be run within your database.

3.2.8 ONNX Format

ONNX (Open Neural Network Exchange) is an open standard format for representing machine learning and deep-learning models. ONNX standardizes the representation and interchange of machine learning models across frameworks, enabling seamless deployment and interoperability. See ONNX for more details.

Select AI can use generative AI models exported in ONNX format to run AI workloads directly inside Autonomous AI Database or through supported runtimes enabling organizations to leverage pre-trained models for natural language processing tasks like query generation. By using ONNX models, you keep inference close to your data, reduce data movement, and enable consistent model processing across different tools and environments ensuring compliant AI operations.

3.2.9 ONNX Runtime

ONNX Runtime runs ONNX-formatted models efficiently across hardware platforms, optimizing inference for real-time AI applications.

Select AI users can specify in-database ONNX-format models in their AI profile in support of RAG. The database embeds the ONNX Runtime in Oracle AI Database 26ai and Autonomous AI Database. Using the in-database ONNX Runtime avoids sending content to an external engine to produce, for example, vector embeddings. ONNX Runtime powers the runtime evaluation of transformer-based models within Autonomous AI Database, facilitating developers to load ONNX models, fast natural language to SQL (NL2SQL) conversions, compute embeddings, classify data, or run inference inside the database engine without sending data to external service, which enhances query performance and improves security, latency, and governance. See Example: Select AI with In-database Transformer Models and ONNX Runtime for more details.

3.2.10 Semantic Similarity Search

Semantic similarity search identifies and retrieves data points that closely match a given query by comparing feature vectors in a vector store.

3.2.11 Similarity Threshold

A similarity threshold sets a minimum score to classify two items as related, filtering results based on their vector proximity or distance. In Select AI, the similarity threshold helps filter results that fall below a required level of semantic closeness, ensuring that only highly related document chunks, rows, or embeddings are returned.