Use Case Artificial Intelligence

Vector Space search has long been employed for Geospatial problems of finding the nearest neighbor over several dimensions where nearest node is not necessarily the closest in the x, y or z dimension, but a combination of all dimensions. It was described as the post-office problem by Donald Knuth in ‘The Art of Computer Programming’ more than 50 years ago , but until recently was mostly used for network routers or statistical segmentation for related products. The most common use of Vector Search is now within the domain of “Artificial Intelligence” where Retrieval-Augmented Generation (RAG) is being used to reduce Hallucination of Large Language Models.

RAG is used to retrieve related information for a ChatGPT prompt where training data may have changed or missing. The process uses the parsed prompt to provide a context vector for searching for all related information, which is then passed to the large language for inference. The LLM ignores unrelated information added as a prelude for the AI prompt. Using very large vectors of subject information can be used for Vector-Space search of a database to retrieve related content.

Hiperspace provided Nearest search in Vector Space using Tensor {Distance, Cosine and Dot product} calculation for match related information in a multi-dimensional space. Hiperspace performs this search/ranking on the raw vector data before it is transformed into business elements. When the vector-space data is stored separately, near to the search servers there is low contention and very fast scanning. Hiperspace uses Aspects to store the vector-data separately to the main information for fast search and update (the aspect contains only the key and (compressed) vector) and can be duplicated for parallel search. “What causes most customer complaints, and has it changed recently?” could identify information-exchange with suppliers if the knowledge graph of subjects is augmented with {process, products, application, service, resource} information.

Hiperspace combines vector-space search with fast retrieval of related information for point-in-time or historical query.

Hiperspace has been designed to take advantage of the new memory technologies being built for scale. Hiperspace was designed for generations of historical data that can be replicated for parallel search.

Hiperspace uses relatively simple key/value stores like RocksDB that are specifically designed for efficient storage in shared memory pools and SSD. In a performance test of Hiperspace against SQL/Server 2019, Hiperspace used 20% of the space and took 20% of the time to search.