|
| 1 | +import aiohttp |
| 2 | + |
| 3 | +from astrbot import logger |
| 4 | + |
| 5 | +from ..entities import ProviderType, RerankResult |
| 6 | +from ..provider import RerankProvider |
| 7 | +from ..register import register_provider_adapter |
| 8 | + |
| 9 | + |
| 10 | +@register_provider_adapter( |
| 11 | + "nvidia_rerank", "NVIDIA Rerank 适配器", provider_type=ProviderType.RERANK |
| 12 | +) |
| 13 | +class NvidiaRerankProvider(RerankProvider): |
| 14 | + def __init__(self, provider_config: dict, provider_settings: dict) -> None: |
| 15 | + super().__init__(provider_config, provider_settings) |
| 16 | + self.api_key = provider_config.get("nvidia_rerank_api_key", "") |
| 17 | + self.base_url = provider_config.get( |
| 18 | + "nvidia_rerank_api_base", "https://ai.api.nvidia.com/v1/retrieval" |
| 19 | + ).rstrip("/") |
| 20 | + self.timeout = provider_config.get("timeout", 20) |
| 21 | + self.model = provider_config.get( |
| 22 | + "nvidia_rerank_model", "nv-rerank-qa-mistral-4b:1" |
| 23 | + ) |
| 24 | + self.model_endpoint = provider_config.get( |
| 25 | + "nvidia_rerank_model_endpoint", "/reranking" |
| 26 | + ) |
| 27 | + self.truncate = provider_config.get("nvidia_rerank_truncate", "") |
| 28 | + |
| 29 | + self.client = None |
| 30 | + self.set_model(self.model) |
| 31 | + |
| 32 | + async def _get_client(self): |
| 33 | + if self.client is None or self.client.closed: |
| 34 | + headers = { |
| 35 | + "Authorization": f"Bearer {self.api_key}", |
| 36 | + "Content-Type": "application/json", |
| 37 | + "Accept": "application/json", |
| 38 | + } |
| 39 | + self.client = aiohttp.ClientSession( |
| 40 | + headers=headers, timeout=aiohttp.ClientTimeout(total=self.timeout) |
| 41 | + ) |
| 42 | + return self.client |
| 43 | + |
| 44 | + def _get_endpoint(self) -> str: |
| 45 | + """ |
| 46 | + 构建完整API URL。 |
| 47 | +
|
| 48 | + 根据 Nvidia Rerank API 文档来看，当前URL存在不同模型格式不一致的问题。 |
| 49 | + 这里针对模型名做一个基础判断用以适配，后续要等Nvidia统一API格式后再做调整。 |
| 50 | +
|
| 51 | + 例： |
| 52 | + 模型： nv-rerank-qa-mistral-4b:1 |
| 53 | + URL: .../v1/retrieval/nvidia/reranking |
| 54 | +
|
| 55 | + 模型： nvidia/llama-nemotron-rerank-1b-v2 |
| 56 | + URL: .../v1/retrieval/nvidia/llama-nemotron-rerank-1b-v2/reranking |
| 57 | + """ |
| 58 | + |
| 59 | + model_path = "nvidia" |
| 60 | + logger.debug(f"[NVIDIA Rerank] Building endpoint for model: {self.model}") |
| 61 | + if "/" in self.model: |
| 62 | + """遵循NVIDIA API的URL规则，替换模型名中特殊字符""" |
| 63 | + model_path = self.model.strip("/").replace(".", "_") |
| 64 | + endpoint = self.model_endpoint.lstrip("/") |
| 65 | + return f"{self.base_url}/{model_path}/{endpoint}" |
| 66 | + |
| 67 | + def _build_payload(self, query: str, documents: list[str]) -> dict: |
| 68 | + """构建请求载荷""" |
| 69 | + payload = { |
| 70 | + "model": self.model, |
| 71 | + "query": {"text": query}, |
| 72 | + "passages": [{"text": doc} for doc in documents], |
| 73 | + } |
| 74 | + if self.truncate: |
| 75 | + payload["truncate"] = self.truncate |
| 76 | + return payload |
| 77 | + |
| 78 | + def _parse_results( |
| 79 | + self, response_data: dict, top_n: int | None |
| 80 | + ) -> list[RerankResult]: |
| 81 | + """解析响应数据""" |
| 82 | + results = response_data.get("rankings", []) |
| 83 | + if not results: |
| 84 | + logger.warning(f"[NVIDIA Rerank] Empty response: {response_data}") |
| 85 | + return [] |
| 86 | + |
| 87 | + rerank_results = [] |
| 88 | + for idx, item in enumerate(results): |
| 89 | + try: |
| 90 | + index = item.get("index", idx) |
| 91 | + score = item.get("relevance_score", item.get("logit", 0.0)) |
| 92 | + rerank_results.append( |
| 93 | + RerankResult(index=index, relevance_score=float(score)) |
| 94 | + ) |
| 95 | + except Exception as e: |
| 96 | + logger.warning( |
| 97 | + f"[NVIDIA Rerank] Result parsing error: {e}, Data={item}" |
| 98 | + ) |
| 99 | + |
| 100 | + rerank_results.sort(key=lambda x: x.relevance_score, reverse=True) |
| 101 | + |
| 102 | + if top_n is not None and top_n > 0: |
| 103 | + return rerank_results[:top_n] |
| 104 | + return rerank_results |
| 105 | + |
| 106 | + def _log_usage(self, data: dict) -> None: |
| 107 | + usage = data.get("usage", {}) |
| 108 | + total_tokens = usage.get("total_tokens", 0) |
| 109 | + if total_tokens > 0: |
| 110 | + logger.debug(f"[NVIDIA Rerank] Token Usage: {total_tokens}") |
| 111 | + |
| 112 | + async def rerank( |
| 113 | + self, |
| 114 | + query: str, |
| 115 | + documents: list[str], |
| 116 | + top_n: int | None = None, |
| 117 | + ) -> list[RerankResult]: |
| 118 | + client = await self._get_client() |
| 119 | + if not client or client.closed: |
| 120 | + logger.error("[NVIDIA Rerank] Client session not initialized or closed") |
| 121 | + return [] |
| 122 | + |
| 123 | + if not documents or not query.strip(): |
| 124 | + logger.warning( |
| 125 | + "[NVIDIA Rerank] Input data is invalid, query or documents are empty" |
| 126 | + ) |
| 127 | + return [] |
| 128 | + |
| 129 | + try: |
| 130 | + payload = self._build_payload(query, documents) |
| 131 | + request_url = self._get_endpoint() |
| 132 | + |
| 133 | + async with client.post(request_url, json=payload) as response: |
| 134 | + if response.status != 200: |
| 135 | + try: |
| 136 | + response_data = await response.json() |
| 137 | + error_detail = response_data.get( |
| 138 | + "detail", response_data.get("message", "Unknown Error") |
| 139 | + ) |
| 140 | + |
| 141 | + except Exception: |
| 142 | + error_detail = await response.text() |
| 143 | + response_data = {"message": error_detail} |
| 144 | + |
| 145 | + logger.error(f"[NVIDIA Rerank] API Error Response: {response_data}") |
| 146 | + raise Exception(f"HTTP {response.status} - {error_detail}") |
| 147 | + |
| 148 | + response_data = await response.json() |
| 149 | + logger.debug(f"[NVIDIA Rerank] API Response: {response_data}") |
| 150 | + results = self._parse_results(response_data, top_n) |
| 151 | + self._log_usage(response_data) |
| 152 | + return results |
| 153 | + |
| 154 | + except aiohttp.ClientError as e: |
| 155 | + logger.error(f"[NVIDIA Rerank] Network error: {e}") |
| 156 | + raise Exception(f"Network error: {e}") from e |
| 157 | + except Exception as e: |
| 158 | + logger.error(f"[NVIDIA Rerank] Error: {e}") |
| 159 | + raise Exception(f"Rerank error: {e}") from e |
| 160 | + |
| 161 | + async def terminate(self) -> None: |
| 162 | + if self.client and not self.client.closed: |
| 163 | + await self.client.close() |
| 164 | + self.client = None |
0 commit comments