Hybrid Search

Hybrid Search is offered by default for "text" and "multimodal" indexes and ensures that your users receive the most relevant results by balancing exact keyword matches with semantic (neural) understanding of broader concepts. Check out API Reference docs for input on how to create and search them.

Features

• Lexical and Semantic Search: Combines exact keyword matching with semantic understanding to deliver comprehensive search results.
• Partial Indexing Support: Objects partially indexed (status: "INCOMPLETE") can still appear in search results if they meet certain criteria.

Limitations

• Partial Indexing: Objects with incomplete indexing status ("INCOMPLETE") can appear in search results if they have at least one successfully indexed component (text embedding, image embedding, or lexical field) but ranking accuracy can be negatively impacted.
• Error Handling: Any errors during indexing are surfaced in the object errors. Retrieve these using the Index Status API or via the Debugging tab in your Console.

FAQs

When should I use Hybrid indexes vs. Neural-only indexes?

• Hybrid Indexes (Default): Choose Hybrid indexes when you need both exact keyword matching and semantic understanding. They are ideal for searches involving specific terms like product names or SKUs, as well as broader concepts.
• Neural-Only Indexes ("text-neural" or "multimodal-neural"): Opt for neural-only indexes when you want to focus solely on semantic relationships without exact keyword matches. These are best for broad use cases like categorization or matching content to topics.

Do you offer lexical-only index configurations?

No, we do not currently offer lexical-only index configurations. If this is something you're interested in, please contact us at [email protected].

Do Hybrid and Neural-only indexes share feature parity?

Yes. All search functionality available for Hybrid indexes is also available for Neural-only ("text-neural" or "multimodal-neural") indexes.

How does the latency of a Hybrid index compare with a Neural-only indexes?

They are comparable in latency.

Do Hybrid Indexes support finetuning?

Yes. You can Finetune a Hybrid index just as you would a text-neural or multimodal-neural index. Behind the scenes we are updating your embedding model but the lexical model is staying the same.

How do objects with INCOMPLETE status impact search results?

Objects with INCOMPLETE status can still appear in search results based on the components that were successfully indexed (text embedding, image embedding, or lexical field). However, their relevance may be inaccurately assessed due to missing data, potentially skewing their ranking in search results. To ensure optimal search performance, it's important to resolve indexing errors so that objects achieve the READY status.