fix: parse default dimension from model metadata string

[BeamNawapat]

Summary

Fix dimension detection for variable-dimension models. Previously hardcoded 1024 for all models with string-format dimensions. Now parses the actual default from the string.

Motivation

Models like voyage-4-nano have a default dimension of 512, not 1024. The string format "512 (default), 128, 256" contains the correct default but was ignored in favor of a hardcoded 1024, causing schema mismatches when creating collections.

Changes

• packages/core/src/embedding/voyageai-embedding.ts — Parse leading number from dimension string with regex instead of hardcoding 1024

Test plan

• pnpm build passes
• Models with 1024 default still return 1024
• Models with 512 default (e.g., voyage-4-nano) correctly return 512

[Copilot]

Pull request overview

This PR updates how VoyageAIEmbedding determines the embedding vector dimension when a model’s dimension metadata is represented as a string (variable-dimension models), avoiding a previously hardcoded default.

Changes:

• Parse the default dimension from the leading numeric token in modelInfo.dimension when it is a string.
• Keep a fallback to 1024 if parsing fails.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The PR description/test plan mentions fixing default dimensions for models like voyage-4-nano (512 default), but this file’s getSupportedModels() does not include that model (and all current string-form dimensions start with 1024). As a result, voyage-4-nano will still hit the unknown-model fallback and keep using 1024, so the reported schema mismatch would persist. Consider adding the missing model(s) to getSupportedModels() (with the correct dimension string) or introducing a safer fallback for unknown models (e.g., detect dimension from an embed response length).

[zc277584121]

Good catch. Parsing the default dimension from the string format (e.g., "1024 (default), 256, 512, 2048") is more robust than hardcoding 1024, especially now that voyage-4-nano has a different default (512).

LGTM.