Issue with the input_type parameter in the NVIDIA plugin for Dify's OpenAI-API-compatible plugin

Discussion
1

Issue: Adding nvidia/nv-embedqa-e5-v5 via the OpenAI-API-compatible plugin results in a 400 error; changing the model name to nv-embedqa-e5-v5 results in a 404 error.

Kimi’s core issue confirmation: The OpenAI-API-compatible plugin in Dify hardcodes the /retrieval path, whereas the correct endpoint for NVIDIA NIM Embedding is /embeddings. This is a 底层 compatibility issue that cannot be resolved via configuration.

Reviewing the historical merge: fix: nvidia special embedding model payload close #11193 by yihong0618 · Pull Request #11239 · langgenius/dify · GitHub

Discovered that the input_type parameter was not being passed. Traced the source code using Claude Code and found the answer:

Problem Analysis

Architecture Layer:

Dify main codebase (api/) ↓ passes input_type
dify-plugin-sdks (OAICompatEmbeddingModel) ← Problem here!
↓ input_type not passed into payload
OpenAI-API-compatible plugin
↓
NVIDIA NIM API (requires input_type parameter)

Root Cause: Although the OAICompatEmbeddingModel._invoke() method in dify-plugin-sdks accepts the input_type parameter, it does not include it when constructing the HTTP payload.

Solution

You need to modify the dify-plugin-sdks repository (not the main Dify repository or dify-official-plugins).

Modify file: python/dify_plugin/interfaces/model/openai_compatible/text_embedding.py

Change the payload construction logic from:

  payload = {
      "input": inputs[i : i + max_chunks],
      "model": credentials.get("endpoint_model_name", model),
      **extra_model_kwargs,
  }

to something like:

  payload = {
      "input": inputs[i : i + max_chunks],
      "model": credentials.get("endpoint_model_name", model),
      **extra_model_kwargs,
  }

  # For models requiring input_type (e.g., NVIDIA NIM)
  if input_type:
      # Convert to format expected by API
      payload["input_type"] = "query" if input_type == EmbeddingInputType.QUERY else "passage"

Final Note: The issue has been confirmed to stem from incorrect parameter passing in the plugin, preventing the addition of NVIDIA’s embedding models.