fix: use input_type instead of task for NVIDIA NIM embeddings

[vicjayjay]

Summary

• NVIDIA NIM API now strictly validates request bodies and rejects unknown fields
• The task parameter causes a 400 error on NVIDIA endpoints
• Fix routes by provider: NVIDIA endpoints get input_type, others get task

Root Cause

buildPayload() in embedder.ts was sending both task and input_type to all providers. NVIDIA rejects task as unknown.

Test plan

• Verified with nvidia/llama-3.2-nemoretriever-1b-vlm-embed-v1
• Non-NVIDIA providers still receive task parameter as before

[rwmjhb]

This is pointing at a real compatibility problem, but I don’t think the current patch is complete enough to merge yet.

Two gaps need to be addressed:

1. input_type value mapping is still wrong for existing config

   • This repo currently documents/configures taskQuery / taskPassage as retrieval.query / retrieval.passage
   • NVIDIA NIM expects input_type values like query / passage, not retrieval.query / retrieval.passage
   • So the current change appears to rename the field without translating the value, which likely still breaks existing NVIDIA configs

2. Provider detection is too narrow

   • The current branch only switches behavior when baseURL matches nvidia.com
   • That misses self-hosted NIM, private gateways, reverse proxies, and other non-nvidia.com deployments
   • If the goal is “NVIDIA NIM compatibility”, URL substring matching alone is not robust enough

Please also add automated coverage for this.
At minimum, I want tests that prove:

• NVIDIA-compatible requests send input_type
• non-NVIDIA providers still send task
• retrieval.query / retrieval.passage are mapped correctly for NVIDIA if that is the intended compatibility path

Once those pieces are in place, I’m happy to take another look.

[rwmjhb]

感谢提交！NVIDIA NIM 拒绝 task 字段的问题是真实存在的。

不过这个 PR 基于的代码已经过时了 — #216 合并后，buildPayload 中的 if (task) payload.task = task; 已被替换为 provider profile 体系：

if (this._capabilities.taskField && task) {
  const value = cap.taskValueMap?.[task] ?? task;
  payload[cap.taskField] = value;
}

现在正确的做法是在 detectEmbeddingProviderProfile 中加一个 "nvidia" profile，然后在 getEmbeddingCapabilities 中设置 taskField: "input_type"，与 Voyage 的处理方式一致。这样不需要在 buildPayload 中硬编码 URL 判断。

如果你有兴趣，欢迎基于最新 master 重新提交。