Working with vector similarity in Neptune Analytics

You can answer complex questions about your data by transforming data shapes into embeddings (that is, vectors). Using a vector search index lets you answer questions about the your data's context and its similarity and connection to other data.

For example, a support agent could translate a question that they receive into a vector and use it to search the support knowledge base for articles that are similar to the words in the question (implicit similarity). For the most applicable articles, they could then collect metadata about the author, previous cases, runbooks, and so on so as to provide additional context when answering the question (explicit data).

Vector similarity search in Neptune Analytics makes it easy for you to build machine learning (ML) augmented search experiences and generative artificial intelligence (GenAI) applications. It also gives you an overall lower total cost of ownership and simpler management overhead because you no longer need to manage separate data stores, build pipelines, or worry about keep the data stores in sync. You can use vector similarity search in Neptune Analytics to augment your LLMs by integrating graph queries for domain-specific context with the results from low-latency, nearest-neighbor similarity search on embeddings imported from LLMs hosted in Amazon Bedrock, Graph Neural Networks (GNNs) in GraphStorm, or other sources.

As an example, Bioinformatics researchers who are interested in re-purposing existing blood pressure drugs for other treatable diseases, want to use vector similarity search over in-house knowledge graphs to find patterns in protein interaction networks.

For another example, a large online book retailer may need to use known pirated material to quickly identify similar media in conjunction with a knowledge graph to identify patterns of deceptive listing behaviours and find malicious sellers.

In both cases, vector search over a knowledge graph increases accuracy and speed when building the solution. It reduces the operational overhead and complexity using the tools available today.

You can create a vector index for your graph to try out this feature. Neptune Analytics supports associating embeddings generated from LLMs with the nodes of your graphs.