Cohere v4 & Qwen3 VL 235B

[BrainWGX]

Could you add support for the Bedrock Multimodel embedding model (like cohere v4) and vision language model (like Qwen3 VL 235B)?

[cscvenkatmadurai]

I've implemented Cohere Embed v4 support in Roo Code — an AI coding assistant that uses Bedrock for code indexing embeddings.

Changes Made

1. Cohere Embed v4 (cohere.embed-v4:0) support:

• Added v4-specific request format with embedding_types: ["float"] parameter
• Added v4-specific response parsing for the {embeddings: {float: [[...]]}} structure (v4 nests embeddings under the type key, unlike v3's flat {embeddings: [[...]]})
• Default dimension: 1536, text-only for now (multimodal image support planned)

2. Fixed credential refresh for long-running indexing sessions:

• For large codebases, indexing can take >1 hour. Temporary AWS credentials (from SSO, STS, or credential_process) would expire mid-session, causing indexing failures
• Replaced fromEnv() with fromNodeProviderChain() as the default credential fallback — supports all credential sources (env vars, SSO, INI, IMDS, ECS) with built-in memoization and auto-refresh

Key Technical Findings from AWS SDK v3 Research

• AWS SDK v3 fromIni() has no built-in memoization for credential_process — it re-runs the subprocess on every invocation. The SDK client itself handles credential caching internally
• ignoreCache: true on fromIni() is irrelevant for credential_process-based profiles — the subprocess is never cached by fromIni() regardless
• The SDK's fromNodeProviderChain() wraps providers with memoize() for efficient caching + auto-refresh near expiry

Implementation: ~10 lines production code across 2 files, 5 new unit tests. All 143 related tests pass.