400 on /api/embed: db_schema embedding receives asyncio Task repr instead of chunk text (Ollama embedder)

[Laharlsama]

Summary

Model deploy fails during semantics preparation with 400 Bad Request on http://.../api/embed. The root cause is that the embedding pipeline passes an asyncio Task object (or its string representation) as the chunk input to the embedder instead of the actual chunk text. Ollama's /api/embed endpoint rejects this invalid input and returns 400.

Environment

• Wren AI version: WREN_AI_SERVICE_VERSION=0.29.0, WREN_UI_VERSION=0.32.2, WREN_ENGINE_VERSION=0.22.0
• Deployment: Self-hosted via Docker Compose on ARM64 (DGX Spark)
• Embedder: Ollama nomic-embed-text:latest (api_base: Ollama container on port 11434)
• LLM: Ollama qwen2.5-coder:7b
• Data source: MySQL; MDL includes multiple tables and many relationships between them

When it appears

• Deploy succeeded earlier with a simpler MDL (fewer or no relationships).
• After defining all relationships between tables in the model, deploy started failing every time.
• So the bug appears when the schema is more complex (many relationships → more chunks / different async scheduling in the pipeline).

Steps to reproduce

1. Create an MDL with several tables and many relationships (e.g. 20+ tables, 30+ relationships).
2. Configure Wren AI with Ollama as default embedder (e.g. ollama/nomic-embed-text:latest) and Ollama as default LLM.
3. Click Deploy on the model.
4. Semantics preparation starts, then fails; in wren-ai-service logs you see 400 on /api/embed and the Node inputs show chunk as a Task repr.

Expected vs actual

• Expected: The pipeline passes chunk text (schema/DDL content) to the embedder. Ollama returns 200 and indexing continues.
• Actual: The pipeline passes a string representation of an asyncio Task (e.g. `"<Task finished name='Task-1479' coro=<AsyncGraphAd...") as the chunk. Ollama returns 400 Bad Request. Deploy fails with "Failed to prepare semantics".

Relevant log excerpt

********************************************************************************
> embedding [src.pipelines.indexing.db_schema.embedding()] encountered an error<
> Node inputs:
{'chunk': "<Task finished name='Task-1479' coro=<AsyncGraphAd...",
 'embedder': '<src.providers.embedder.litellm.AsyncDocumentEmbed...'}
********************************************************************************
Traceback (most recent call last):
  File "/app/.venv/lib/python3.12/site-packages/litellm/main.py", line 3882, in aembedding
    response = await init_response  # type: ignore
  ...
  File "/app/.venv/lib/python3.12/site-packages/httpx/_models.py", line 763, in raise_for_status
    raise HTTPStatusError(message, request=request, response=self)
httpx.HTTPStatusError: Client error '400 Bad Request' for url 'http://kitsune_ollama:11434/api/embed'
...
  File "/src/pipelines/indexing/db_schema.py", line 318, in embedding
    return await embedder.run(documents=chunk["documents"])
...
E0228 13:21:03.276 7 wren-ai-service:100] Failed to prepare semantics: litellm.APIConnectionError: OllamaException - Client error '400 Bad Request' for url 'http://kitsune_ollama:11434/api/embed'

So the input to the embedding node is wrong: chunk is the string repr of a Task, not the actual documents/chunk text. This likely comes from an upstream step in the Hamilton graph returning or passing a Task without awaiting it.

Config (relevant parts only)

Embedder (default):

  - alias: default
    api_base: http://kitsune_ollama:11434
    model: ollama/nomic-embed-text:latest
    timeout: 300
    kwargs:
      num_gpu: -1

LLM (default):

  - alias: default
    api_base: http://kitsune_ollama:11434
    model: ollama_chat/qwen2.5-coder:7b
    timeout: 900

Additional context

• Ollama is reachable and works for chat and for embeddings when called directly with valid text.
• The failure is reproducible with the same MDL every time.
• Workaround used so far: none that fixes the root cause; reducing model complexity (fewer relationships) might avoid the failing code path but is not a real fix.

[hirebbec]

support