Bug Description

Summary

When configuring OpenViking with provider: jina and model: jina-code-embeddings-1.5b, the server fails at startup with a 422 Validation Error from the Jina API. The root cause is that OpenViking internally passes task: "retrieval.query" to the Jina embeddings endpoint, but jina-code-embeddings-1.5b is a code-specialized model that only accepts code-specific task types.

Environment

• OpenViking version: latest (pip)
• OS: macOS (Apple M4, arm64)
• Python: 3.12.11 (pyenv)
• Jina model: jina-code-embeddings-1.5b

Expected Behavior

OpenViking should either:

1. Auto-detect the correct task type based on the model name (e.g. when model is jina-code-embeddings-*, use nl2code.query / nl2code.passage as defaults)
2. Expose a config field (e.g. task_type) so users can explicitly set the task prefix in ov.conf

Suggested Fix

In embedding_config.py or jina_embedders.py, add model-aware task mapping:

CODE_EMBEDDING_MODELS = {"jina-code-embeddings-1.5b", "jina-code-embeddings-0.5b"}

def get_task_type(model: str, role: str) -> str:
    if model in CODE_EMBEDDING_MODELS:
        return f"nl2code.{role}"  # role = "query" or "passage"
    return f"retrieval.{role}"

Steps to Reproduce

Config (ov.conf)

{
  "embedding": {
    "dense": {
      "provider": "jina",
      "model": "jina-code-embeddings-1.5b",
      "dimension": 1024
    }
  }
}

Error Message

openviking: agent memories search failed: Error: OpenViking request failed [INTERNAL]:
Jina API error: Error code: 422 - {
  'detail': "Validation error: 'body -> jina-code-embeddings-1.5b -> task'
  Input should be 'nl2code.query', 'nl2code.passage', 'qa.query', 'qa.passage',
  'code2code.query', 'code2code.passage', 'code2nl.query', 'code2nl.passage',
  'code2completion.query' or 'code2completion.passage'.",
  ...
  'errors': [{
    'field': 'body -> jina-code-embeddings-1.5b -> task',
    'message': "Input should be 'nl2code.query' ...",
    'type': 'literal_error',
    'input': 'retrieval.query'
  }]
}

Root Cause

The Jina embedder implementation in jina_embedders.py appears to hardcode a generic task type (retrieval.query / retrieval.passage) for all Jina models. However, jina-code-embeddings-1.5b is a code-specialized model that uses a different task namespace:

Expected Behavior

OpenViking should either:

1. Auto-detect the correct task type based on the model name (e.g. when model is jina-code-embeddings-*, use nl2code.query / nl2code.passage as defaults)
2. Expose a config field (e.g. task_type) so users can explicitly set the task prefix in ov.conf

Suggested Fix

In embedding_config.py or jina_embedders.py, add model-aware task mapping:

CODE_EMBEDDING_MODELS = {"jina-code-embeddings-1.5b", "jina-code-embeddings-0.5b"}

def get_task_type(model: str, role: str) -> str:
    if model in CODE_EMBEDDING_MODELS:
        return f"nl2code.{role}"  # role = "query" or "passage"
    return f"retrieval.{role}"

Workaround

Switch to jina-embeddings-v3 which supports the generic retrieval.query task type. However, this sacrifices the significantly better code retrieval quality of jina-code-embeddings-1.5b.

Happy to submit a PR if the fix direction is confirmed. Thanks!

Actual Behavior

Error Message

openviking: agent memories search failed: Error: OpenViking request failed [INTERNAL]:
Jina API error: Error code: 422 - {
  'detail': "Validation error: 'body -> jina-code-embeddings-1.5b -> task'
  Input should be 'nl2code.query', 'nl2code.passage', 'qa.query', 'qa.passage',
  'code2code.query', 'code2code.passage', 'code2nl.query', 'code2nl.passage',
  'code2completion.query' or 'code2completion.passage'.",
  ...
  'errors': [{
    'field': 'body -> jina-code-embeddings-1.5b -> task',
    'message': "Input should be 'nl2code.query' ...",
    'type': 'literal_error',
    'input': 'retrieval.query'
  }]
}

Root Cause

The Jina embedder implementation in jina_embedders.py appears to hardcode a generic task type (retrieval.query / retrieval.passage) for all Jina models. However, jina-code-embeddings-1.5b is a code-specialized model that uses a different task namespace:

Workaround

Switch to jina-embeddings-v3 which supports the generic retrieval.query task type. However, this sacrifices the significantly better code retrieval quality of jina-code-embeddings-1.5b.

Happy to submit a PR if the fix direction is confirmed. Thanks!

Minimal Reproducible Example

Error Logs

### Error Message


openviking: agent memories search failed: Error: OpenViking request failed [INTERNAL]:
Jina API error: Error code: 422 - {
  'detail': "Validation error: 'body -> jina-code-embeddings-1.5b -> task'
  Input should be 'nl2code.query', 'nl2code.passage', 'qa.query', 'qa.passage',
  'code2code.query', 'code2code.passage', 'code2nl.query', 'code2nl.passage',
  'code2completion.query' or 'code2completion.passage'.",
  ...
  'errors': [{
    'field': 'body -> jina-code-embeddings-1.5b -> task',
    'message': "Input should be 'nl2code.query' ...",
    'type': 'literal_error',
    'input': 'retrieval.query'
  }]
}

OpenViking Version

openviking 0.2.6

Python Version

Python 3.12.11

Operating System

macOS

Model Backend

Other

Additional Context

No response