Voyage AI: Voyage Code 2

Voyage Code 2 creates numerical embeddings for source code and related text. It is designed for semantic code retrieval, where a developer describes behavior or supplies a code fragment and the search system finds relevant implementations. The maker’s examples cover natural-language questions, code snippets and combinations of the two as search inputs.

Use it to build a repository search experience around actual development questions. A developer might look for the routine that removes only the first and last matching characters, or for the part of a project that applies a particular tree traversal. Prepare queries that express the behavior precisely, including the condition that distinguishes the intended implementation from a nearby but incorrect one.

Index code in useful units rather than splitting every file at an arbitrary character count. Keep a function signature, body and relevant explanation together. When a helper depends on local types or constants, retain enough surrounding text to make its purpose understandable. Store file paths, repository revisions and source locations outside the embedding so the result can open the original code directly.

This edition has a 16,000-token context window and produces 1,536-dimensional embeddings. Size the vector index for this representation and keep query and document embeddings associated with the same model. A collection already indexed with Voyage Code 2 can retain a consistent retrieval representation while its source records are updated. A move to another embedding edition should be treated as an index migration with an explicit comparison on the same repository.

For a natural-language query, describe the operation, inputs and relevant outcome. For a code query, include the fragment whose logic needs to be matched and explain any condition that matters to the search. A search for retry handling, for example, can specify whether the relevant code retries after a timeout or after a complete error response. The retrieved source should be checked for that distinction before it is used in an edit.

Keep documentation and source code traceable when both are indexed. An explanatory passage may be helpful context, while the implementation is the artifact a developer needs to inspect. Label these record types so a search page can show the correct destination and a coding assistant can distinguish a description from executable source.

Use the retrieved passages as context for repository navigation, implementation discovery or a coding assistant’s next step. Preserve the complete relevant snippet and its location when passing it onward. Similarity helps find candidate code; reviewing the actual inputs, side effects and contracts determines whether the candidate solves the current problem.

Evaluate retrieval on a set of real questions with known source answers. Include renamed functions, misleading comments, similar utilities with different edge cases and queries that use vocabulary absent from the code. Check whether the expected implementation appears among the useful results and whether the displayed snippet gives enough context to assess it. Retain those questions as the repository grows so changes to extraction, passage boundaries or indexing can be compared against the same tasks.