Integrating voyage-2 into an application typically involves calling its embedding API and wiring the results into your data pipeline. At a minimum, you need a small wrapper function that sends text to the API and receives embeddings as arrays of floats. This function can be used both in offline jobs (for indexing documents) and in online request handlers (for embedding user queries). Because the API is stateless, integration is usually straightforward from most backend environments.
A common architecture looks like this: an ingestion service reads source data, chunks it, calls voyage-2 in batches, and stores the embeddings. An application service handles user queries, embeds the query text, and performs similarity search. The two services can be decoupled, which makes scaling easier. For example, ingestion might run as a scheduled batch job, while query embedding happens in real time. Error handling, retries, and rate limiting are typical concerns, but they are standard API-integration problems rather than ML-specific challenges.
For storage and retrieval, most applications integrate voyage-2 with a vector database such as Milvus or Zilliz Cloud. These databases provide client SDKs and APIs that make it easy to insert vectors, build indexes, and query for nearest neighbors. The application code then simply combines the query embedding with a vector search call. This integration pattern keeps the system modular: voyage-2 handles embedding generation, the database handles similarity search, and the application handles business logic.
For more information, click here: https://zilliz.com/ai-models/voyage-2