Hybrid Search using Hybrid Vector Indexes

You can use a hybrid vector index to index and query data using a combination of full-text search and vector similarity search. Hybrid searches can enhance the relevance (quality) of your search results by integrating the keyword matching capabilities of Oracle Text indexes with the semantic precision of vector indexes.

An end-to-end indexing pipeline facilitates the indexing of documents without requiring you to be an expert in various text processing, chunking, or embedding strategies.

As part of this feature, new preference helper procedures have been added to the existing DBMS_VECTOR and DBMS_VECTOR_CHAIN PL/SQL packages to manage vector index-related settings.

A new PL/SQL package, DBMS_HYBRID_VECTOR, has also been added. The package contains a SEARCH API that allows you to query against hybrid vector indexes in multiple ways.

A new dictionary view <index name>$VECTORS provides information on row identifiers, chunks, and embeddings for all indexed documents.

• Manage Hybrid Vector Indexes

• Perform Hybrid Search

• DBMS_HYBRID_VECTOR

• CREATE_PREFERENCE

• DROP_PREFERENCE

• <index name>$VECTORS

Updated AI Vector Search Workflow

The AI Vector Search Workflow included in this guide has been updated to include support for hybrid vector indexes and hybrid search.

Oracle AI Vector Search Workflow

The documentation map to GenAI prompts has been updated to include the new features available in Release Update 23.6.

Your Vector Documentation Map to GenAI Prompts

The SQL Quick Start Using a Vector Embedding Model Uploaded into the Database has been extended to include steps for creating a hybrid vector index and running a hybrid search.

SQL Quick Start Using a Vector Embedding Model Uploaded into the Database

Support for Ollama

You can use open embedding models (such as mxbai-embed-large, nomic-embed-text, or all-minilm) and large language models (such as Llama 3, Phi 3, Mistral, or Gemma 2) using the local host REST endpoint provider, Ollama. Ollama is a free and open-source command-line interface tool that allows you to run these models locally and privately on your Linux, Windows, and macOS systems.

• Convert Text String to Embedding Using the Local REST Provider Ollama
• Generate Summary Using the Local REST Provider Ollama
• Generate Text Using the Local REST Provider Ollama

Support for SPARSE Vectors

In addition to DENSE vectors, you have the option to create SPARSE vectors. With SPARSE vectors, dimension values are mostly zero and only the non-zero values are physically stored.

The COMPATIBLE initialization parameter must be set to 23.6.0 to use this feature.

Create Tables Using the VECTOR Data Type

Integration with LlamaIndex

Oracle AI Vector Search's integration with LlamaIndex allows you to use LlamaIndex's open-source data framework with vector search capabilities. This provides a powerful foundation for building sophisticated AI applications that can leverage both structured and unstructured data within the Oracle ecosystem.

Oracle AI Vector Search Integration with LlamaIndex

Jaccard Distance

JACCARD is available as a distance metric and JACCARD_DISTANCE as a shorthand for the VECTOR_DISTANCE function to calculate the Jaccard distance between two BINARY vectors.

• Jaccard Similarity

• JACCARD_DISTANCE

Hamming Distance

HAMMING_DISTANCE is available as a shorthand for the VECTOR_DISTANCE function to calculate the Hamming similarity between two vectors.

HAMMING_DISTANCE

Transaction Support for HNSW Indexes

Transactions are maintained on tables with Hierarchical Navigable Small World (HNSW) index graphs by using data structures, such as private journal and shared journal.

A private journal records vectors that are added or deleted by a transaction, whereas a shared journal records all the commit system change numbers (SCNs) and corresponding modified rows.

If using RAC instances, you must enable Patch 36932885 using Oracle RAC two-stage rolling updates.

Understand Transaction Support for Tables with HNSW Indexes

HNSW Index Duplication and Reload

By default, a duplication mechanism is used to create HNSW indexes in an Oracle RAC environment or when an HNSW index repopulation operation is triggered. When an Oracle Database instance starts again, a reload mechanism is triggered to recreate the HNSW graph in memory as quickly as possible. HNSW full checkpoints are used to reload the graph in memory.

You can use the VECTOR_INDEX_NEIGHBOR_GRAPH_RELOAD initialization parameter to manage duplication and reload mechanisms for HNSW indexes.

If using RAC instances, you must enable Patch 36932885 using Oracle RAC two-stage rolling updates.

Understand HNSW Index Population Mechanisms in Oracle RAC or Single Instance

Update to the default for VECTOR_INDEX_NEIGHBOR_GRAPH_RELOAD

The default value for the VECTOR_INDEX_NEIGHBOR_GRAPH_RELOAD initialization parameter has been changed from OFF to RESTART. With this update, HNSW duplication and reload mechanisms are enabled by default.

Understand HNSW Index Population Mechanisms in Oracle RAC or Single Instance

Global and Local Partitioning for IVF Indexes

Inverted File Flat (IVF) vector indexes support both global and local indexes on partitioned tables.

By default, IVF indexes are globally partitioned by centroid. You can choose to create a local IVF index, which provides a one-to-one relationship between the base table partitions or subpartitions and the index partitions.

If using RAC instances, you must enable Patch 36932885 using Oracle RAC two-stage rolling updates.

• Inverted File Flat Vector Indexes Partitioning Schemes

• Inverted File Flat Index Syntax and Parameters

GET_INDEX_STATUS, ENABLE_CHECKPOINT, DISABLE_CHECKPOINT, and INDEX_VECTOR_MEMORY_ADVISOR Procedures

New procedures are available with the DBMS_VECTOR PL/SQL package:

• GET_INDEX_STATUS to query the status of a vector index creation

• ENABLE_CHECKPOINT and DISABLE_CHECKPOINT to enable or disable the Checkpoint feature for HNSW indexes

• INDEX_VECTOR_MEMORY_ADVISOR to determine the vector memory size needed for a vector index

Vector Index Status, Checkpoint, and Advisor Procedures

VECSYS.VECTOR$INDEX$CHECKPOINTS View

A new dictionary view VECSYS.VECTOR$INDEX$CHECKPOINTS provides information about HNSW full checkpoints at the database level.

VECSYS.VECTOR$INDEX$CHECKPOINTS

Relational Data Vectorization

You can use Oracle Machine Learning (OML) Feature Extraction algorithms with the VECTOR_EMBEDDING() SQL operator to vectorize sets of relational data, build similarity indexes, and perform similarity searches. These algorithms help in extracting the most informative features or columns from the data, making it easier to analyze correlations and redundancies in that data.

Vectorize Relational Tables Using OML Feature Extraction Algorithms

You can use the RERANK function, added to the existing DBMS_VECTOR and DBMS_VECTOR_CHAIN PL/SQL packages, to enhance the relevancy of your query results in Retrieval Augmented Generation (RAG) scenarios. Third-party reranking models are used to reassess and reorder an initial list of retrieved documents.

• Use Reranking for Better RAG Results

• DBMS_VECTOR.RERANK

• DBMS_VECTOR_CHAIN.RERANK

BLOB Support for UTL_TO_GENERATE_TEXT()

In addition to the existing textual input, the UTL_TO_GENERATE_TEXT() PL/SQL function accepts BLOB as input. You can provide binary data, such as an image file, to generate a textual analysis or meaningful description of the contents of an image.

Describe Images Using Public REST Providers

BLOB Support for UTL_TO_EMBEDDING()

In addition to the existing textual input, the UTL_TO_EMBEDDING() PL/SQL function accepts BLOB as input. This lets you directly generate a vector embedding from image files using REST API calls to third-party image embedding models.

Convert Image to Embedding Using Public REST Providers

List of REST Endpoints

You can refer to the list of REST endpoints and corresponding REST calls supported for all third-party REST providers.

Supported Third-Party Provider Operations and Endpoints