Models from an external OpenAI-compatible connection are missing from /api/models when prefix_id is empty

Bug Description

When an external OpenAI-compatible connection has an empty prefix_id, the models it returns are correctly fetched when querying that connection individually (/openai/models/{idx}), but they are silently dropped when the models are merged/aggregated into the final list at /api/models. No error is logged.

Adding a non-empty prefix_id to the same connection immediately fixes the issue and the models appear in /api/models.

Environment

• open-webui image: ghcr.io/open-webui/open-webui:main
• Deployment: Docker, self-hosted
• Backend proxy: LiteLLM (ghcr.io/berriai/litellm:main-latest), acting as an OpenAI-compatible endpoint on the same Docker network

Steps to Reproduce

1. Add an external OpenAI API connection pointing to an OpenAI-compatible endpoint (in my case, a local LiteLLM proxy at http://litellm:4000/v1), with prefix_id left empty.
2. Enable the connection (enable: true), verify connection succeeds (POST /openai/verify returns 200).
3. Query that connection's models directly:
   GET /openai/models/{connection_index}
   → Returns the full list of models correctly (confirmed 8 models returned, e.g. packy.aws-q/claude-sonnet-5, packy.cc-sale/claude-sonnet-5, etc.)
4. Query the aggregated model list:
   GET /api/models
   → The models from this connection are completely absent, even though other connections (with non-empty prefix_id) show up correctly.
5. Edit the connection and set prefix_id to any non-empty string (e.g. lite/).
6. Query /api/models again → the models now appear correctly (prefixed as expected).

Expected Behavior

Models from a connection with an empty prefix_id should still appear in the aggregated /api/models response, the same as models from a connection with a prefix_id set — only the model ID prefixing should differ, not whether the models are included at all.

Actual Behavior

Models are silently dropped during aggregation when prefix_id is an empty string (""). No error, no warning in logs — connection verification and individual model fetch (/openai/models/{idx}) both succeed.

Additional Context

• OPENAI_API_CONFIGS for the affected connection (via GET /openai/config):

  {
      "enable": true,
      "tags": [],
      "prefix_id": "",
      "model_ids": [],
      "connection_type": "external",
      "auth_type": "bearer",
      "passthrough_params": []
  }

• Other connections in the same config, with non-empty prefix_id, work as expected.
• Logs during the failing request show only 200 OK responses for all relevant endpoints (/openai/verify, /openai/models/{idx}, /api/models) — no exceptions or errors are logged, making this hard to diagnose.

Suggested Fix Direction

The model-merging logic likely uses prefix_id as part of a key (e.g., for deduplication or grouping), and an empty string may be causing models to be overwritten, filtered, or excluded unintentionally during that step. Would appreciate a look at whatever function assembles the aggregated model list at /api/models to see how it treats connections with an empty/falsy prefix_id.

Workaround

Setting a non-empty prefix_id (e.g. a short string like lite/) on the affected connection resolves the issue completely.

Start by tracing the aggregation behind GET /api/models and compare it with GET /openai/models/{connection_index}, focusing on how an empty prefix_id is handled during merging or deduplication. Reproduce with an enabled external connection whose prefix_id is empty, then confirm its models appear in the aggregated response without a prefix while existing prefixed connections still work.