Voyage AI: Voyage Context 3

Voyage Context 3 produces contextualized chunk embeddings for document retrieval. Each chunk receives its own vector, while the model also considers the other chunks supplied from the same document. This gives a retrieval system a representation of the local passage together with broader document context, such as an entity or topic introduced elsewhere in the source.

Prepare each document as an ordered sequence of readable text chunks. Keep the chunks from one document together and preserve their original order. In the maker’s contextualized workflow, each inner group contains the chunks of one source, and separate groups represent separate sources. This grouping is part of the input’s meaning: it tells the model which surrounding text belongs to the document being represented.

A useful case is a report whose opening identifies the company while a later paragraph explains its growth. A search for that company’s growth needs both the specific paragraph and the document identity. Keep the original chunk text and its parent-document metadata with the resulting vector so the result can display the passage and link to the complete source.

The maker’s reference documents up to 120,000 tokens across a request, with no more than 1,000 input groups and 16,000 chunks in total. The model guide separately lists a 32,000-token window per chunk. Keep these scopes distinct when preparing material: a limit on a single chunk and a limit on the complete request describe different input sizes. Check the total after dividing the source rather than assuming every individual chunk can fill its maximum in the same request.

For document retrieval, identify the grouped passages as document input. Prepare each search question as its own query group. The maker recommends non-overlapping supplied document chunks, allowing the contextualized representation to carry surrounding information without repeating text in every passage. Retain the correspondence between document group, chunk position and returned vector when writing the index.

Voyage Context 3 supports 2,048, 1,024, 512 and 256 dimensions, with floating-point, eight-bit integer and binary output choices. Choose an index representation and evaluate it against the same source documents and questions. Keep the selected dimension and data type with the index so the query representation and stored vectors remain consistent.

Use the resulting vectors in passage search, document discovery or the retrieval stage of an assistant working with reports, research papers, contracts or technical manuals. A result can identify a focused passage while its source link gives access to definitions and qualifications elsewhere in the document. Preserve document versions and permissions as application metadata alongside the vectors.

Test questions that require both a local detail and document-level context. Include similar passages from different sources, named entities introduced only at the beginning and questions whose evidence spans several sections. Inspect whether the expected chunk is retrieved and whether the application presents enough surrounding text to interpret it. When a source changes, regenerate the affected document group deliberately so its contextualized chunks correspond to one coherent version.