Oracle AI Vector Search

Similarity search is focused on finding related data based on its semantic meaning. Unstructured data is difficult to search directly, so similarity search goes beyond simple keyword searches by considering the underlying text, image, audio, or video data instead of only searching the labels manually applied to it.

The need to identify a match for similar data across large data sets applies to many industries. Examples of similarity search include the following:

• Semantic search of document
• Semantic search of images
• Recommendation systems
• Pattern recognition
• Anomaly and fraud detection

1. Unstructured data such as files, documents, BLOBS or CLOBS.
2. Use commercial or open source embedding models to create either vectors inside or outside the Oracle Database.
3. The vectors can be stored next to the business data such that when the data changes the vectors are also changing.
4. You can search the stored vectors for the most similar vectors.
5. The search returns the most relevant data for you.

RAG uses the results of similarity search to improve the accuracy and contextual relevance of large language model responses to questions about business data. RAG helps identify contextually relevant private data that the LLM may not have been trained on and then uses it to augment user prompts so LLMs can respond with greater accuracy.

The desire to get higher quality answers from LLMs is universal, spanning many industries. Some examples of using RAG for improved accuracy include the following:

• Chatbots for internal and external users
• Document searches and summaries
• Language to code synthesis
• Answers to questions that require specialized, domain-specific knowledge

RAG helps organizations provide customized answers to business questions without the high cost of retraining or fine-tuning the LLMs.

1. A chatbot enables a dialog with an LLM.
2. Run similarity search on your private business data and pass those facts to the LLM.
3. The results are formatted as a prompt and context for the LLM.
4. The LMM receives up to date business data inputs thereby reducing hallucinations.
5. The high-quality responses are returned to the chatbot.