wooo/awoooi
1
1
Fork 0
Code Issues 2 Pull Requests 1 Actions 2 Packages Projects Releases Wiki Activity

feat(ai): ADR-036 NVIDIA Nemotron Tool Calling 整合

Browse Source 
Phase 20 - 提升 Tool Calling 精準度 50% → 83.3%

新增:
- src/models/nvidia.py: Pydantic Schema
- src/services/nvidia_provider.py: NvidiaProvider 類別
- tests/test_nvidia_provider.py: 15 項單元測試 (全部通過)

修改:
- ai_router.py: AIProvider.NVIDIA + route_tool_calling()
- ai_rate_limiter.py: NVIDIA 限制 (5 RPM, 100/day)
- models.json: NVIDIA 配置
- cd.yaml: Secrets 注入 NVIDIA_API_KEY

路由策略:
- Tool Calling: Nemotron → Gemini → Claude
- 一般對話: Ollama → Gemini → Claude (不變)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
OG T
2026-03-29 00:00:08 +08:00
parent dc7daf5d81
commit b77e151387
11 changed files with 1083 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
apps/api/src/services/ai_router.py
View File

@@ -66,6 +66,8 @@ class AIProvider(Enum):
OLLAMA = "ollama"
GEMINI = "gemini"
CLAUDE = "claude"
# 2026-03-29 ogt: ADR-036 Nemotron Tool Calling (83.3% 精準度)
NVIDIA = "nvidia"
# Provider 對應延遲預算 (ms)
@@ -73,6 +75,8 @@ PROVIDER_LATENCY_BUDGET: dict[AIProvider, int] = {
AIProvider.OLLAMA: 60000, # 本地，允許較長處理時間
AIProvider.GEMINI: 30000, # 雲端，較低延遲
AIProvider.CLAUDE: 30000, # 雲端，較低延遲
# 2026-03-29 ogt: ADR-036 Nemotron Tool Calling (延遲 11-45s)
AIProvider.NVIDIA: 60000, # Tool Calling 專用，允許較長時間
}
@@ -164,21 +168,32 @@ class AIRouter:
self._ollama_summary = self._model_registry.get_model("ollama", "summary")
self._gemini_default = self._model_registry.get_model("gemini", "default")
self._claude_default = self._model_registry.get_model("claude", "default")
# 2026-03-29 ogt: ADR-036 Nemotron Tool Calling
self._nvidia_default = self._model_registry.get_model("nvidia", "default")
# Provider 對應模型映射
self._provider_models: dict[AIProvider, str] = {
AIProvider.OLLAMA: self._ollama_default,
AIProvider.GEMINI: self._gemini_default,
AIProvider.CLAUDE: self._claude_default,
AIProvider.NVIDIA: self._nvidia_default, # ADR-036
}
# 完整 Fallback 鏈 (Provider, Model)
# 2026-03-29 ogt: NVIDIA 不在一般 Fallback 鏈 (僅用於 Tool Calling)
self._full_fallback_chain: list[tuple[AIProvider, str]] = [
(AIProvider.OLLAMA, self._ollama_default),
(AIProvider.GEMINI, self._gemini_default),
(AIProvider.CLAUDE, self._claude_default),
]
# Tool Calling 專用 Fallback 鏈 (ADR-036)
self._tool_calling_fallback_chain: list[tuple[AIProvider, str]] = [
(AIProvider.NVIDIA, self._nvidia_default),
(AIProvider.GEMINI, self._gemini_default),
(AIProvider.CLAUDE, self._claude_default),
]
# 意圖對應 Provider 強制覆寫 (None = 依複雜度決定)
self._intent_provider_overrides: dict[IntentType, AIProvider | None] = {
# 四大核心意圖
@@ -466,6 +481,39 @@ class AIRouter:
routing_latency_ms=routing_latency,
)
# =========================================================================
# Tool Calling 路由 (ADR-036)
# =========================================================================
def route_tool_calling(self) -> tuple[AIProvider, str, list[tuple[AIProvider, str]]]:
"""
Tool Calling 專用路由 (ADR-036)
Tool Calling 任務優先使用 Nemotron (83.3% 精準度)，
Fallback 到 Gemini/Claude。
Returns:
(provider, model, fallback_chain)
"""
provider = AIProvider.NVIDIA
model = self._nvidia_default
fallback_chain = [
(p, m) for p, m in self._tool_calling_fallback_chain if p != provider
]
logger.info(
"tool_calling_routing",
provider=provider.value,
model=model,
fallback_count=len(fallback_chain),
)
return provider, model, fallback_chain
def get_tool_calling_fallback_chain(self) -> list[tuple[AIProvider, str]]:
"""取得 Tool Calling Fallback 鏈"""
return self._tool_calling_fallback_chain.copy()
# =========================================================================
# 便捷方法
# =========================================================================