VoyageAI by MongoDB: voyage-code-4

voyage-code-4 is Voyage AI’s code-specialised embedding model for repository search and coding-agent retrieval. It is designed to help locate relevant source when the starting point is a question, a symptom or a task description rather than a known identifier. It represents code and textual information as numerical vectors for retrieval; it does not edit the repository or generate a fix.

Retrieve code from the issue being investigated

The maker describes a training corpus curated from completed pull requests, connecting natural-language queries with code. This is a meaningful distinction from training examples built mainly around source files and their docstrings: the retrieval task can start with the behavior that needs investigation.

Prepare a query such as “Find where an empty result causes a successful operation to be reported as a failure.” Index readable source sections with their file paths and relevant comments. Keep the function’s surrounding conditions where they explain its behavior. A returned candidate should lead the reader or agent back to the source section it represents.

Semantic retrieval can complement exact identifier and full-text search. If a function name is already known, an exact lookup remains useful. If only a symptom is known, vector search can provide candidates to inspect and refine with another query.

Query and document roles

Voyage’s retrieval guidance distinguishes the query from the corpus document through query/document input types. For code search, use the issue or task description as the query and the indexed source or documentation as the document. The published recipe adds role-specific retrieval instructions during embedding.

Keep the code’s original formatting, names and relevant documentation available outside the vector. For a source section, retain its file path and location; for a programming guide, retain its heading. After retrieval, inspect the actual code and its callers before choosing an edit. A candidate’s relevance does not establish that it is the sole cause of a reported problem.

Context and representation choices

The exact model’s published context length is 32,000 tokens. Its default embedding has 1024 dimensions, with 256, 512 and 2048 dimensions also supported through Matryoshka learning. The announcement documents 32-bit floating-point, signed and unsigned 8-bit integer, and binary representations.

Select one dimension and representation for the collection, and keep them with its metadata. Floating-point vectors and quantized representations require the corresponding storage and comparison handling. A binary representation encodes vector information differently from a list of ordinary floating-point coordinates; preserve that distinction when preparing results for search.

For a long file, choose coherent sections whose source locations can be recovered. A function may need an adjacent type definition or comment to explain a condition, while a whole oversized file may include unrelated material. Keep the section aligned with the task to be searched rather than treating the maximum context as a preferred chunk size.

Use retrieved source as evidence

A practical coding-agent workflow retrieves candidate source, reads those files, follows relevant references and narrows the search as the investigation progresses. The retrieved text becomes context for the reasoning or editing stage. Keep its association with the indexed repository revision so that a candidate is inspected against the correct source.

voyage-code-4’s edition-specific contribution is its focus on the issue-to-code retrieval step used by coding agents, together with flexible dimensions and representations. Its output is a retrieval vector, while diagnosis, editing and validation remain separate tasks grounded in the actual repository.