Summary

• add a dedicated VoyageDenseEmbedder for Voyage's OpenAI-compatible embeddings API
• register voyage as a first-class dense embedding provider in config
• use Voyage model defaults when the embedding schema dimension is omitted
• document the supported Voyage configuration and add focused unit coverage

Why this version is intentionally narrow

This PR adds Voyage dense text embeddings in a way that matches OpenViking's current embedding architecture.

Today OpenViking:

• configures a single dense embedder for both indexing and query-time retrieval
• stores and validates dense vectors as float vectors with a fixed schema dimension

Because of that, this PR intentionally does not expose Voyage-specific features that do not fit the current architecture yet:

• no output_dtype config field
• no query/document mode in config

Instead, the PR keeps the public surface to the existing generic fields:

• provider
• api_key
• api_base
• model
• dimension

For Voyage, dimension is mapped internally to Voyage's output_dimension request field.

Review feedback addressed

This revision directly addresses the review concerns from #614:

1. output_dtype is not added to EmbeddingModelConfig

   • OpenViking currently expects dense float vectors across storage and retrieval.
   • Exposing Voyage quantized output types now would add provider-specific config surface without a matching end-to-end runtime path.

2. query/document mode is not documented as a supported config knob

   • OpenViking currently uses one dense embedder configuration for both indexing and querying.
   • Documenting input_type: document/query at this stage would imply asymmetric retrieval support that the current config model does not provide yet.

3. Voyage support still remains first-class

   • the provider is explicit in config and validation
   • the implementation is isolated in a dedicated embedder instead of being hidden inside the OpenAI path

Implementation details

• openviking/models/embedder/voyage_embedders.py
  • adds VoyageDenseEmbedder
  • validates allowed output dimensions for Voyage models that support dimension control
  • uses Voyage model defaults when no dimension is specified
• openviking_cli/utils/config/embedding_config.py
  • validates provider: \"voyage\"
  • returns Voyage model default dimensions through get_effective_dimension() when schema dimension is omitted
  • wires the new provider into the dense embedder factory
• docs/en/guides/01-configuration.md
  • adds a Voyage config example
  • documents the current architectural limits clearly
• tests
  • add unit coverage for config validation, dimension defaults, request shaping, and error handling

Verification

• .venv/bin/python -m pytest tests/unit/test_voyage_embedder.py tests/unit/test_embedding_config_voyage.py --noconftest -o addopts='' -q
• .venv/bin/python -m pytest tests/misc/test_config_validation.py -o addopts='' -q
• .venv/bin/ruff check openviking/models/embedder/voyage_embedders.py openviking_cli/utils/config/embedding_config.py tests/unit/test_embedding_config_voyage.py tests/unit/test_voyage_embedder.py