Environment

• OpenViking: v0.2.9 (pip install)
• Python: 3.12.3
• Ollama: local, with nomic-embed-text (768-dim) and llama3.2:3b
• OS: Ubuntu 24.04

Bug 1: inf similarity scores when using Ollama embedding provider

Reproduction

1. Configure ov.conf with Ollama dense embeddings (without explicit dimension):

{
  "embedding": {
    "dense": {
      "provider": "ollama",
      "model": "nomic-embed-text",
      "api_base": "http://localhost:11434/v1",
      "api_key": "ollama"
    }
  }
}

2. Start server, add a resource, wait for indexing
3. Search: POST /api/v1/search/find → {"status":"error","error":{"message":"Out of range float values are not JSON compliant: inf"}}

Root Cause

In openviking_cli/utils/config/embedding_config.py, get_effective_dimension() has no special case for the ollama provider — it falls through to return 2048. But nomic-embed-text returns 768-dim vectors.

This causes:

1. Collection created with Dim: 2048
2. Ollama returns 768-dim vectors → dimension mismatch errors during initial indexing
3. If dimension: 768 is later added to config, new vectors are stored as 768-dim but the collection index still expects 2048
4. At search time, cosine similarity between mismatched dimensions → float('inf') → JSON serialization error

Suggested Fix

Add Ollama dimension auto-detection in get_effective_dimension():

if provider == "ollama":
    try:
        from openviking.models.embedder.openai_embedders import OpenAIDenseEmbedder
        embedder = OpenAIDenseEmbedder(
            model_name=self.model or "nomic-embed-text",
            api_key=self.api_key or "no-key",
            api_base=self.api_base or "http://localhost:11434/v1",
            dimension=None,  # triggers _detect_dimension()
        )
        return embedder.get_dimension()
    except Exception:
        pass

Alternatively, document that dimension is required when using the Ollama provider.

Bug 2: VLM api_base not forwarded to OpenAI client

Reproduction

1. Configure VLM with Ollama:

{
  "vlm": {
    "model": "llama3.2:3b",
    "api_base": "http://localhost:11434/v1",
    "api_key": "ollama"
  }
}

2. Trigger memory extraction (session commit)
3. Error: 401 - Incorrect API key provided: ollama — the request goes to api.openai.com instead of localhost:11434

Root Cause

In openviking_cli/utils/config/vlm_config.py, _build_vlm_config_dict() only forwards api_key/api_base from a providers dict entry (lines 163-166). But _migrate_legacy_config() only populates the providers dict when both self.api_key and self.provider are set. Since the simple VLM config doesn't include a provider field, get_provider_config() returns (None, None), and api_base is never forwarded to the OpenAI client.

Suggested Fix

Add a fallback in _build_vlm_config_dict():

if config:
    result["api_key"] = config.get("api_key")
    result["api_base"] = config.get("api_base")
    result["extra_headers"] = config.get("extra_headers")
else:
    # Fall back to top-level fields when no provider config found
    if self.api_key:
        result["api_key"] = self.api_key
    if self.api_base:
        result["api_base"] = self.api_base
    if self.extra_headers:
        result["extra_headers"] = self.extra_headers

Workarounds

1. Embedding dimension: Always set dimension explicitly in ov.conf for the Ollama provider
2. VLM api_base: Set OPENAI_API_KEY and OPENAI_BASE_URL environment variables as a workaround, or add a provider field to the VLM config

Both bugs were discovered while integrating OpenViking v0.2.9 with a local Ollama instance for an AI agent memory system. Happy to submit PRs for either fix.