Semantic Search

A demo of AI semantic search where a query is matched by context and meaning, rather than keyword.

About

Semantic Search lets you search across a set of documents using a query based on context and meaning, rather than keyword matching.

The app works by using the Cohere large language model (LLM) to retrieve embeddings for the query and the list of items in the database. A cosine similiarity score is calculated for each database item against the query. The best match is highlighted.

The app uses artificial intelligence with an LLM to calculate embeddings. An embedding is a numerical representation of a piece of text and includes many features including meaning, context, size, color, shape, topic, etc. Each feature is converted into a floating point value to create a vector.

Example of an embedding

the cat and the fox => [ 0.12, 0.55, 0.76, 0.11, 0.87, .. ]

Similarity is then be calculated for the embeddings by comparing the distance between the query vector and each database item vector.

Screenshot

Quick Start

Install the required libraries in VSCode before running the project.

Open a Terminal in VSCode.

Create a file .env with the contents:

CohereApiKey=<YOUR_API_KEY>
OpenAIApiKey=<YOUR_API_KEY>

You can register a Cohere API key here.

You can register an OpenAI API key here.

Deployment

To build a deployment for hosting on a web server, use the following steps.

In Visual Studio Code, run the Publish build step.
Copy the .env file into /publish and add your API key for Cohere.
Run the executable /pubish/SemanticSearch.exe

The web application will begin running on port 5000. The executable is a self-contained web server, similar to a node.js application, and may be deployed to a Windows web server or Microsoft Azure.

Saving Embeddings

Applications should save embeddings to a database to prevent the need to call the LLM for documents that have already been processed. Ideally, this should be a vector database, which is designed to store embeddings and perform built-in similarity calculations. However, to keep this project simple, embeddings are always retrieved from the LLM.

See the branch database for saving the embeddings to a local database as a cache.