Oracle Cloud Infrastructure Documentation

Select AI with Retrieval Augmented Generation (RAG)

Select AI with RAG augments your natural language prompt by retrieving content from your specified vector store using semantic similarity search. This reduces hallucinations by using your specific and up-to-date content and provides more relevant natural language responses to your prompts.

Select AI automates the Retrieval Augmented Generation (RAG) process. This technique retrieves data from enterprise sources using AI vector search and augments user prompts for your specified large language model (LLM). By leveraging information from enterprise data stores, RAG reduces hallucinations and generates grounded responses.

RAG uses AI vector search on a vector index to find semantically similar data for the specified question. Vector store processes vector embeddings, which are mathematical representations of various data points like text, images, and audio. These embeddings capture the meaning of the data, enabling efficient processing and analysis. For more details on vector embeddings and AI vector search, see Overview of AI Vector Search.

Select AI integrates with AI vector search available in Oracle Autonomous Database 23ai for similarity search using vector embeddings.

Topics

Parent topic: Use Select AI for Natural Language Interaction with your Database

Build your Vector Store

Select AI automates the creation and population of vector store by converting input documents (for example, PDF, DOC, JSON, XML, or HTML) from your object store to plain text. Oracle Text supports around 150 file types. For a complete list of all the supported document formats, see Supported Document Formats.

Select AI automatically processes documents to chunks, generates embeddings, stores them in the specified vector store, and updates the vector index as new data arrives.

Here is how the input from Object Storage is used with Select AI RAG:
Description of adb-obj-storage-select-ai-rag.png follows
Description of the illustration adb-obj-storage-select-ai-rag.png

  1. Input: Data is initially stored in an Object Storage.
  2. Oracle Autonomous Database retrieves the input data or the document, chunks it and sends the chunks to an embedding model.
  3. The embedding model processes the chunk data and returns vector embeddings.
  4. The vector embeddings are then stored in a vector store for use with RAG. As content is added, the vector index is automatically updated.

RAG retrieves relevant pieces of information from the enterprise database to answer a user's question. This information is provided to the specified large language model along with the user prompt. Select AI uses this additional enterprise information to enhance the prompt, improving the LLM's response. RAG can enhance response quality with update-to-date enterprise information from the vector store.


Description of adb-diagram-third-party-vector-db-select-ai-rag.eps follows
Description of the illustration adb-diagram-third-party-vector-db-select-ai-rag.eps

Select AI implements RAG as follows:

  1. Input: User asks a question (specifies a prompt) using Select AI narrate action.

  2. Select AI generates vector embeddings of the prompt using the embedding model specified in the AI profile.

  3. The vector search index uses the vector embedding of the question to find matching content from the customer's enterprise data (searching the vector store) which has been indexed.

  4. The vector search returns top K texts similar to the input to your Autonomous Database instance.
  5. Autonomous Database then sends these top K query results with user question to the LLM.
  6. The LLM returns its response to your Autonomous Database instance.
  7. Autonomous Database Select AI provides the response to the user.

Use DBMS_CLOUD_AI to Create and Manage Vector Indexes

Use the DBMS_CLOUD_AI package to create and manage vector indexes and configure vector database JSON parameters.

After you create credentials and provide network access to the vector database and the AI provider, your Autonomous Database instance uses AI profiles to configure access to LLMs. See Example: Set Up and Use Select AI with RAG for a complete example on setting it up and using it in Select AI statements.
Note

If you do not want table data or vector search documents to be sent to an LLM, a user with administrator privileges can disable such access for all users of the given database. This, in effect, disables the narrate action for RAG.

You can configure AI profiles for providers listed in Select your AI Provider and LLMs through the DBMS_CLOUD_AI package.

See Also:

Parent topic: Select AI with Retrieval Augmented Generation (RAG)

Use In-database Transformer Models

Select AI RAG enables you to use pretrained ONNX transformer models that are imported into your database in Oracle Database 23ai instance for generating embedding vectors from document chunks and user prompts.

To learn more about importing a pretrained ONNX model into your Oracle Database 23ai instance, see
Note

You must import a pretrained ONNX-format transformer model into Oracle Database 23ai instance to use Select AI RAG with imported in-database transformer model. You can also use other transformer models from supported AI providers.

See Example: Select AI with In-database Transformer Models to explore the feature.

Benefits of Select AI RAG

Simplify querying, enhance response accuracy with current data, and gain transparency by reviewing sources used by the LLM.

Select AI RAG offers the following benefits:

  • Simplify data querying and increase response accuracy: Enable users to query enterprise data using natural language and provide LLMs with detailed context from enterprise data to generate more accurate and relevant responses, reducing instances of LLM hallucinations.

  • Up-to-date information: Provide LLMs access to current enterprise information using vector stores, eliminating the need for costly, time-consuming fine-tuning of LLMs trained on static data sets.

  • Seamless integration: Integrate with Oracle AI Vector Search for streamlined data handling and enhanced performance.

  • Automated data orchestration: Automate orchestration steps with a fully managed Vector Index pipeline, ensuring efficient processing of new data.

  • Understandable contextual results: Has access and retrieves the sources used by the LLM from vector stores, ensuring transparency and confidence in results. Views and extracts data in natural language text or JSON format for easier integration and application development.