feat(voyage): add voyage-context-4 and voyage-4 family models

[fzowl]

Summary

Adds VoyageAI's latest models and fills in missing pricing/context data.

New models

• voyage/voyage-context-4 — contextualized chunk embeddings (latest), $0.12/M, 120K max total tokens
• voyage/voyage-4 — $0.06/M, 32K context
• voyage/voyage-4-large — $0.12/M, 32K context
• voyage/voyage-4-lite — $0.02/M, 32K context
• voyage/voyage-4-nano — open-weight, 32K context

Added to both model_prices_and_context_window.json and litellm/model_prices_and_context_window_backup.json.

Contextual embeddings input

The Voyage contextualized embeddings API expects inputs as List[str] or List[List[str]]. transform_embedding_request now normalizes a bare str into a List[str] so the API is always called with the shape it accepts; List[str] / List[List[str]] pass through unchanged.

Docs & tests

• Updated docs/my-website/docs/providers/voyage.md with the new models and input formats.
• Added tests in tests/llm_translation/test_voyage_ai.py for context-4 detection and string normalization.

To double-check

• Pricing/context values pulled from Voyage docs (pricing + embeddings + contextualized-chunk pages).
• voyage-4-nano is open-weight with no listed per-token price; included in model map without a price override.

[fzowl]

VERDICT:APPROVED

Reviewed model config, pricing, contextual-input handling, and lineup completeness against the current Voyage docs (embeddings, pricing, contextualized-chunk, multimodal pages). Solid PR — ships correctly.

Pricing / config — all correct

Model	PR value	Voyage docs	✓
voyage-4	$0.06 (6e-08)	$0.06	✓
voyage-4-large	$0.12 (1.2e-07)	$0.12	✓
voyage-4-lite	$0.02 (2e-08)	$0.02	✓
voyage-4-nano	0.0 (open-weight)	no public price	✓
voyage-context-4	$0.12 (1.2e-07), 120K	$0.12, 120K total	✓
voyage-multimodal-3.5	$0.12 (1.2e-07), 32K	$0.12/M text	✓

Context windows consistent with existing siblings (context-4 = 120000 matches context-3; multimodal-3.5 = 32000 matches multimodal-3). Both JSON files updated identically.

Contextual input normalization — correct

API accepts List[List[str]] for pre-chunked docs. The PR normalizes str -> [[input]] and flat List[str] -> [input], passing List[List[str]] through unchanged. Produces a valid API shape; fixes the prior bug where a bare str would be sent as "inputs": "..." (invalid). is_contextualized_embeddings("voyage-context-4") matches via "context" in model. Tests cover all three shapes + detection.

Lineup — complete for embeddings

All current Voyage embedding models present: voyage-4 family, context-4, multimodal-3.5 added; code-3 / 3.5 / context-3 / multimodal-3 / finance-2 / law-2 already existed. No missing embedding models. (Rerank models out of scope.)

Non-blocking notes

1. voyage-multimodal-3.5 is undisclosed. Added to both JSON files but absent from the PR body's "New models" list and from voyage.md docs. Real, current model — but mention it in the description and add a docs row for consistency.
2. Flat List[str] -> single document. Defensible choice, but Voyage treats a flat List[str] as either auto-chunking (enable_auto_chunking=True) or multiple queries. Wrapping as one document's chunks is a reasonable default and yields valid output; worth a one-line doc note so callers know the semantics.
3. voyage-4-nano cost 0.0. Fine given no public price (open-weight, not in Voyage's free-tier list). If it ever gets a served API price, cost tracking would undercount.

None block merge.

[fzowl]

Correct the followings:

• lint fails
• the transform_embedding_request's response should provide inputs as a list[str]! (unless list[list[str]] is received, in that case leave as it is)

[fzowl]

Starting fresh to test the v2 flow end-to-end.