[Bug] ov reindex --regenerate processes Semantic queue but LLM overview content never written to .overview.md

[pty819]

Bug Description

ov reindex --regenerate triggers Semantic queue processing (queue counter increments) but LLM-generated overview/abstract content is never written to .overview.md / .abstract.md files — they remain as [Directory overview is not generated] placeholders.

The Semantic queue consumer runs (processed count increments) but the VLM is never called and no content is written.

---

Version

• Server: OpenViking v0.3.13 (released 2026-04-29, running as openviking-server)
• CLI: ov v0.3.11.dev0 (Python package)
• Embedding model: v5-small-retrieval-Q8_0.gguf (local GGUF, confirmed working — 17,357 vectors in VikingDB)
• VLM model: nvidia_nim/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning via https://integrate.api.nvidia.com/v1
• Python: 3.10+
• OS: Linux

---

Reproduction Steps

1. Add a resource and wait for embeddings to complete

# Add any directory resource
ov add-resource ~/finance_dict/

# Monitor — embeddings complete but LLM overviews are never generated
ov status

2. Trigger reindex with --regenerate

TARGET_URI="viking://resources/finance_dict/What_Are_Zombie_Companies__Understanding__5707d2ba"

ov reindex --regenerate --wait "$TARGET_URI"
# Output: {"status": "success", "enqueued_count": 1}

3. Observe Semantic queue processes items

# Before reindex:
# [queue] Semantic: 0 processed

# After reindex:
# [queue] Embedding: 15377 | Semantic: 4 processed | Semantic-Nodes: 4 processed

4. Check output files — still placeholders

$ cat ~/.ovk_workspace/viking/default/resources/finance_dict/.../.overview.md
# What_Are_Zombie_Companies__Understanding__5707d2ba
[Directory overview is not generated]

$ cat ~/.ovk_workspace/viking/default/resources/finance_dict/.../.abstract.md
[Abstract not generated]

---

Expected Behavior

After ov reindex --regenerate --wait, the .overview.md and .abstract.md files should contain LLM-generated content (L0 abstract ~100 tokens, L1 overview ~2k tokens).

The OpenAPI spec at crates/ov_cli/src/commands/reindex.rs states:

"If regenerate=True, also regenerates L0/L1 summaries via LLM before re-embedding"

---

Actual Behavior

• Semantic queue processes 4 items (increments from 0 → 4)
• .overview.md and .abstract.md are touched (mtime updated to reindex time)
• File contents remain placeholder strings: [Directory overview is not generated] / [Abstract not generated]
• ov status shows VLM model calls are not tracked in the [models] section (only the embedding GGUF model appears)
• No VLM API calls are made (NVIDIA NIM dashboard confirms zero calls for this model during reindex)

---

Root Cause Hypothesis

The semantic queue consumer runs and marks items as processed=4, but appears to skip the VLM generation step. Possible causes:

1. Semantic processor has a guard that skips content writing — the generate_content or overview_type=auto logic may be disabled/broken in v0.3.13
2. Workspace write path mismatch — semantic processor writes to a different workspace path than the one being read
3. VLM client initialization failure in semantic worker — VLM client is initialized for embedding pipeline but not for summarization pipeline in the semantic worker
4. Silent VLM failure — VLM API call returns an error that is swallowed silently by the semantic processor

---

Verification: VLM Works Directly

The NIM nemotron model works correctly when called directly:

curl -s -X POST https://integrate.api.nvidia.com/v1/chat/completions \
  -H "Authorization: Bearer <key>" \
  -H "Content-Type: application/json" \
  -d '{"model":"nvidia/nemotron-3-nano-omni-30b-a3b-reasoning","messages":[{"role":"user","content":"What are zombie companies?"}],"max_tokens":50}'

# Response: valid JSON with coherent text — VLM is functional

Additionally, vikingbot (which uses the same OpenViking server with the same nemotron config) successfully calls the VLM for chat — proving the API key is real and the VLM integration works in the bot pipeline.

---

Queue State Snapshot

[queue] (healthy)
+----------------+---------+-------------+-----------+----------+--------+-------+
|     Queue      | Pending | In Progress | Processed | Requeued | Errors | Total |
+----------------+---------+-------------+-----------+----------+--------+-------+
|   Embedding    |    0    |      0      |   15377   |  11792   |   0    | 15377 |
|    Semantic    |    0    |      0      |     4     |    0     |   0    |   4   |
| Semantic-Nodes |    0    |      0      |     4     |    0     |   0    |   4   |
|     TOTAL      |    0    |      0      |   15381   |  11792   |   0    | 15381 |
+----------------+---------+-------------+-----------+----------+--------+-------+

[vikingdb] (healthy)
+------------+-------------+--------------+--------+
| Collection | Index Count | Vector Count | Status |
+------------+-------------+--------------+--------+
|  context   |      1      |    17357     |   OK   |
+------------+-------------+--------------+--------+

[models] (healthy)
Embedding Models:
+------------------------------+----------+-------+--------+------------+--------+
|            Model             | Provider | Calls | Prompt | Completion | Total  |
+------------------------------+----------+-------+--------+------------+--------+
| v5-small-retrieval-Q8_0.gguf |  openai  | 3614  | 143287 |     0      | 143287 |
+------------------------------+----------+-------+--------+------------+--------+
# NOTE: No VLM model entries — VLM was never called during reindex

---

Additional Context

• This is a regression — embeddings were completed (17,357 vectors) and reindex --regenerate is the intended way to generate LLM overviews for pre-existing resources
• The resource finance_dict contains ~6100 markdown files that all have embeddings but zero LLM overviews
• All .overview.md / .abstract.md files were created as empty placeholders at ingest time

[pty819]

Root Cause Confirmed

TL;DR: ov.conf vlm.model 配置错误，多了 nvidia_nim/ 前缀，导致 NVIDIA NIM chat completions API 返回 404。

根因

ov.conf 里 vlm.model 配置为：

nvidia_nim/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning

实际模型名为：

nvidia/nemotron-3-nano-omni-30b-a3b-reasoning

models.list() 能识别带 nvidia_nim/ 前缀的名称（因为它只做 list），但 chat.completions.create() 的 404 端点不接受这个格式。

验证

# models.list 成功
curl https://integrate.api.nvidia.com/v1/models -H "Authorization: Bearer nvapi-..."
# 返回: nvidia/nemotron-3-nano-omni-30b-a3b-reasoning ✓

# chat.completions 404
curl https://integrate.api.nvidia.com/v1/chat/completions \
  -d '{"model":"nvidia_nim/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",...}'
# 返回: 404 page not found

# 正确模型名仍然 403（Authorization failed）
curl https://integrate.api.nvidia.com/v1/chat/completions \
  -d '{"model":"nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",...}'
# 返回: 403 Authorization failed

但 vikingbot 能正常聊天。这说明 vikingbot 使用了不同的模型路由。

---

重要配置备注：vikingbot 需要 nvidia_nim/ 前缀

vikingbot 使用 litellm 调用 NVIDIA NIM 时，模型名必须带 nvidia_nim/ 前缀，否则 litellm 报错 LLM Provider NOT provided。

即 vikingbot 的模型配置应该是：

nvidia_nim/nvidia/nemotron-3-nano-omni-30b-a3b-reasoning

而 OpenViking Server 的 semantic processor（走 OpenAI SDK）应该用：

nvidia/nemotron-3-nano-omni-30b-a3b-reasoning

这是因为 litellm 和 OpenAI Python 客户端对 NVIDIA NIM 模型的命名解析逻辑不同。

建议修复

在 openai_vlm.py 的 _call 方法里，打印/记录实际的 model 参数，方便用户排查 404 错误。is_available() 通过不代表 chat.completions.create() 能成功 — 两者检查的端点不同。

[pty819]

详细分析报告：OpenViking 模型配置的 consumer 不一致问题

问题现象

在 ov.conf 中配置 vlm.model = "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning" 时：

• OpenViking Server（内容处理 / LLM overview 生成）：✅ 正常工作
• vikingbot CLI (uv run openviking chat): ❌ 报错

  litellm.BadRequestError: LLM Provider NOT provided. 
  Pass in the LLM provider you are trying to call.
  You passed model=nvidia/nemotron-3-nano-omni-30b-a3b-reasoning
  

---

根本原因：两个 consumer 共享同一份配置，但底层 SDK 机制不同

Consumer 1：OpenViking Server（--with-bot）

路径：

ov.conf['vlm'] 
  → VLMFactory.create(config) 
  → provider="openai" → OpenAIVLM (openai Python SDK)
  → self.client.chat.completions.create(model="nvidia/nemotron-...", ...)

关键代码：openai_vlm.py 直接用 self.model（即 ov.conf['vlm']['model']）构建请求：

# openai_vlm.py, _build_text_kwargs()
kwargs = {
    "model": self.model,  # 直接使用 ov.conf 中配置的 model
    "messages": kwargs_messages,
    ...
}
response = client.chat.completions.create(**kwargs)

NVIDIA NIM API 行为：

• 接受 bare model name 格式（nvidia/nemotron-...）
• 拒绝带 nvidia_nim/ 前缀的格式
• models.list() 可以正常列出 nvidia/nemotron-...，但 chat.completions.create() 对 nvidia_nim/nvidia/... 返回 404

结论：Server 端用 nvidia/nemotron-... ✅（已修复）

---

Consumer 2：vikingbot CLI (uv run openviking chat)

路径：

ov.conf 
  → loader._merge_vlm_model_config(bot_data, vlm_data)
  → bot_data['agents']['model'] = vlm_data['model']  
  → LiteLLMProvider(default_model=model)
  → litellm.completion(model=model, ...)

关键代码：loader.py 第 108-120 行：

def _merge_vlm_model_config(bot_data: dict, vlm_data: dict) -> None:
    if vlm_data.get("model"):
        model = vlm_data["model"]  # 读入 "nvidia/nemotron-..."
        provider = vlm_data.get("provider")  # "openai"
        if provider and "/" not in model:
            model = f"{provider}/{model}"  # 不会执行，model 已含 "/"
        bot_data["agents"]["model"] = model  # "nvidia/nemotron-..."

然后在 litellm_provider.py 第 188 行：

model = self._resolve_model(model or self.default_model)
# → detect_provider_by_model("nvidia/nemotron-...") → None
# → api_base 存在但不匹配 openai/hosted_vllm/ollama 前缀
# → litellm 报错 "LLM Provider NOT provided"

关键函数 litellm_vlm.py 的 detect_provider_by_model()：

PROVIDER_CONFIGS = {
    "openrouter": {"keywords": ("openrouter",), ...},
    "anthropic": {"keywords": ("claude", "anthropic"), ...},
    "openai": {"keywords": ("gpt", "o1", "o3", "o4"), ...},
    # 注意：没有 "nvidia" 或 "nvidia_nim" 条目！
}

def detect_provider_by_model(model: str) -> str | None:
    model_lower = model.lower()
    for provider, config in PROVIDER_CONFIGS.items():
        if any(kw in model_lower for kw in config["keywords"]):
            return provider
    return None  # "nvidia/nemotron-..." 匹配不到任何 provider

---

为什么 litellm 需要 nvidia_nim/ 前缀

当 litellm 的 detect_provider_by_model() 返回 None 时：

# litellm_provider.py, _resolve_model()
if provider and provider in PROVIDER_CONFIGS:
    prefix = PROVIDER_CONFIGS[provider]["litellm_prefix"]
    ...
else:
    if self.api_base and not model.startswith(...):
        return f"openai/{model}"  # 变成 "openai/nvidia/nemotron-..."
    return model

api_base = "https://integrate.api.nvidia.com/v1" 存在，但 model 不以 openai//hosted_vllm//ollama/ 开头，所以被强制加上了 openai/ 前缀 → 变成 openai/nvidia/nemotron-... → 传给 NVIDIA NIM → 404。

而 litellm 官方支持的 nvidia_nim/ 前缀（参考 litellm 文档）会将 model 重写为 nvidia_nim/nvidia/nemotron-...，litellm 内部识别 nvidia_nim provider 后，会将请求转发到配置的 api_base，同时保留完整 model 路径。

---

配置矛盾总结

配置字段	Server (OpenAIVLM)	vikingbot CLI (LiteLLMProvider)
vlm.model	nvidia/nemotron-... ✅	nvidia/nemotron-... ❌
实际需要	bare model name	nvidia_nim/nvidia/nemotron-...
SDK 行为	直接发送 model 给 API	检测 provider → 路由失败

---

建议修复方案

核心问题：vlm.model 是单一配置字段，但被两个 consumer（OpenAI SDK vs LiteLLM）以不同方式解释。

建议在 vlm 配置中增加分离字段：

{
  "vlm": {
    "provider": "openai",
    "api_base": "https://integrate.api.nvidia.com/v1",
    "api_key": "nvapi-...",
    "model": "nvidia/nemotron-...",        // 用于 Server (OpenAIVLM)
    "bot_model": "nvidia_nim/nvidia/nemotron-..."  // 用于 vikingbot CLI (LiteLLM)
  }
}

或者在 litellm_vlm.py 的 PROVIDER_CONFIGS 中添加 nvidia_nim 条目，让 LiteLLM 能正确识别 NVIDIA NIM 模型：

"nvidia_nim": {
    "keywords": ("nvidia_nim",),
    "env_key": "NVIDIA_API_KEY",  
    "litellm_prefix": "nvidia_nim",
}

---

环境信息

• OpenViking Server: v0.3.13 (Python 3.13, .venv)
• vikingbot: v0.3.13
• litellm: 内置于 .venv
• NVIDIA NIM endpoint: https://integrate.api.nvidia.com/v1
• nemotron API key: nvapi-ArK5Lwmv_JbZm（已验证有效）
• ov.conf 路径: ~/ovk/ov.conf