Mariangeles Carranza
We’re excited to announce a new integration of Voyage AI into Version 2.1 of the Pureinsights Discovery platform a big step forward in delivering smarter and more context-aware search experiences.
Voyage AI brings powerful tools through its API, offering advanced endpoints such as embeddings, multimodal embeddings, and rerankers. With this integration, we’ve enabled support for these features across both Discovery QueryFlow and Discovery Ingestion modules.
What’s New in the Discovery Platform 2.1?
The integration of Voyage AI has been implemented across two key components of our products (via the Voyage AI Embeddings API):
Discovery Ingestion Voyage AI Processor
- Embeddings Action: Apply semantic vector encoding during ingestion generated data.
- Multimodal Embeddings Action: Apply semantic vector encoding during ingestion generated data, but supporting text and images.
Discovery QueryFlow Voyage AI Processor
- Embeddings Action: Encode text documents into semantic vectors.
- Multimodal Embeddings Action: Combine text documents and images into a shared embedding space.
- Rerankers Action: Reorder documents based on the query.
What Are Embeddings?
Embeddings are vector representations that capture the semantic meaning behind words and phrases, allowing models to recognize similarity even if different words are used.
For example, consider the following two sentences:
- How to bake a cake
- Cake recipes
Even though they use different words, their embeddings will be close in vector space meaning the system understands they are related in meaning.
What Are Multimodal Embeddings?
Multimodal embeddings take things a step further by allowing you to combine text and images into the same semantic space.
Currently, the integration supports images in the following formats:
- .jpg, .jpeg, .png, .gif, and .webp, but provided as base64-encoded strings or image URLs
What Are Rerankers?
Rerankers are another exciting capability of Voyage AI. Once a set of initial search results are returned, rerankers can reorder the results based on how well they match the context behind the user’s query. This significantly improves the relevance of the top results, ensuring that users see the most helpful content first.
A Real-World Example
Imagine you’re building a digital library for recipes, articles, or product documentation, and your users want to search for something.
Here’s how this integration helps:
- With Embeddings: The system understands the meaning behind the query and retrieves related results even if the wording is different.
This means that if asks, “How to bake a cake?”, the system can understand and provide results with some cake recipes, even if they don’t match the exact words.
- With Rerankers: Among the retrieved results, the system reorders the most contextually relevant ones to the top.
- With Multimodal Embeddings: Users can search using both text, images or both. Let’s look at this example of a recipe page:
1. Imagine that we have a collection of recipes, in which each recipe is described with text and images:
2. In Discovery Ingestion, with the multimodal embeddings action of the Voyage AI component we can create the embeddings for the recipes, with text and images in the same vector space. Normally, images and texts are completely different types of data but in this case, the model forces itself to learn a shared “meaning”: similar images and texts end up close together in vector space.
- The multimodal embedding for the user query.
- The recipes index executes a vector search with the user input.
- Similar documents are retrieved.
4. The results are displayed to the user.
As a result, multimodal embeddings improve analysis because they allow meaningful, cross-format matching between text and images, leading to more flexible, accurate, and user-friendly search across different types of data.
Final Thoughts
The integration of Voyage AI into the Pureinsights Discovery Platform unlocks the ability to build smarter search applications. Whether users are typing a question, pasting a block of text, or uploading an image, Discovery can deliver semantically relevant results that go far beyond simple keyword matching.
If you would like more information on Discovery or help with your AI application, please CONTACT US for a free consultation.