|
| 1 | +import aiohttp |
| 2 | + |
| 3 | +from astrbot import logger |
| 4 | + |
| 5 | +from ..entities import ProviderType |
| 6 | +from ..provider import EmbeddingProvider |
| 7 | +from ..register import register_provider_adapter |
| 8 | + |
| 9 | + |
| 10 | +@register_provider_adapter( |
| 11 | + "nvidia_embedding", |
| 12 | + "NVIDIA NIM Embedding 提供商适配器", |
| 13 | + provider_type=ProviderType.EMBEDDING, |
| 14 | +) |
| 15 | +class NvidiaEmbeddingProvider(EmbeddingProvider): |
| 16 | + def __init__(self, provider_config: dict, provider_settings: dict) -> None: |
| 17 | + super().__init__(provider_config, provider_settings) |
| 18 | + self.provider_config = provider_config |
| 19 | + self.provider_settings = provider_settings |
| 20 | + |
| 21 | + self.api_key = provider_config.get("embedding_api_key", "") |
| 22 | + self.base_url = ( |
| 23 | + provider_config.get( |
| 24 | + "embedding_api_base", "https://integrate.api.nvidia.com/v1" |
| 25 | + ) |
| 26 | + .rstrip("/") |
| 27 | + .removesuffix("/embeddings") |
| 28 | + ) |
| 29 | + self.timeout = int(provider_config.get("timeout", 20)) |
| 30 | + self.model = provider_config.get( |
| 31 | + "embedding_model", "nvidia/llama-nemotron-embed-1b-v2" |
| 32 | + ) |
| 33 | + self.input_type = provider_config.get("input_type", "passage") |
| 34 | + |
| 35 | + proxy = provider_config.get("proxy", "") |
| 36 | + self.proxy = proxy |
| 37 | + if proxy: |
| 38 | + logger.info(f"[NVIDIA Embedding] Using proxy: {proxy}") |
| 39 | + |
| 40 | + self.client = None |
| 41 | + self.set_model(self.model) |
| 42 | + |
| 43 | + async def _get_client(self): |
| 44 | + if self.client is None or self.client.closed: |
| 45 | + headers = { |
| 46 | + "Authorization": f"Bearer {self.api_key}", |
| 47 | + "Content-Type": "application/json", |
| 48 | + "Accept": "application/json", |
| 49 | + } |
| 50 | + timeout = aiohttp.ClientTimeout(total=self.timeout) |
| 51 | + self.client = aiohttp.ClientSession( |
| 52 | + headers=headers, |
| 53 | + timeout=timeout, |
| 54 | + ) |
| 55 | + return self.client |
| 56 | + |
| 57 | + def _build_payload(self, text: str | list[str]) -> dict: |
| 58 | + if isinstance(text, str): |
| 59 | + input_text = [text] |
| 60 | + else: |
| 61 | + input_text = text |
| 62 | + |
| 63 | + return { |
| 64 | + "input": input_text, |
| 65 | + "model": self.model, |
| 66 | + "input_type": self.input_type, |
| 67 | + "encoding_format": "float", |
| 68 | + } |
| 69 | + |
| 70 | + def _parse_response(self, response_data: dict) -> list[list[float]]: |
| 71 | + data = response_data.get("data", []) |
| 72 | + embeddings = [] |
| 73 | + for item in data: |
| 74 | + embedding = item.get("embedding", []) |
| 75 | + embeddings.append(embedding) |
| 76 | + return embeddings |
| 77 | + |
| 78 | + async def get_embedding(self, text: str) -> list[float]: |
| 79 | + embeddings = await self.get_embeddings([text]) |
| 80 | + return embeddings[0] if embeddings else [] |
| 81 | + |
| 82 | + async def get_embeddings(self, text: list[str]) -> list[list[float]]: |
| 83 | + client = await self._get_client() |
| 84 | + if not client or client.closed: |
| 85 | + raise Exception("[NVIDIA Embedding] Client session not initialized") |
| 86 | + |
| 87 | + payload = self._build_payload(text) |
| 88 | + request_url = f"{self.base_url}/embeddings" |
| 89 | + |
| 90 | + try: |
| 91 | + async with client.post( |
| 92 | + request_url, json=payload, proxy=self.proxy or None |
| 93 | + ) as response: |
| 94 | + if response.status != 200: |
| 95 | + error_text = await response.text() |
| 96 | + logger.error( |
| 97 | + f"[NVIDIA Embedding] API Error: {response.status} - {error_text}" |
| 98 | + ) |
| 99 | + raise Exception( |
| 100 | + f"NVIDIA Embedding API request failed: HTTP {response.status} - {error_text}" |
| 101 | + ) |
| 102 | + |
| 103 | + response_data = await response.json() |
| 104 | + embeddings = self._parse_response(response_data) |
| 105 | + |
| 106 | + usage = response_data.get("usage", {}) |
| 107 | + total_tokens = usage.get("total_tokens", 0) |
| 108 | + if total_tokens > 0: |
| 109 | + logger.debug(f"[NVIDIA Embedding] Token usage: {total_tokens}") |
| 110 | + |
| 111 | + return embeddings |
| 112 | + |
| 113 | + except aiohttp.ClientError as e: |
| 114 | + logger.error(f"[NVIDIA Embedding] Network error: {e}") |
| 115 | + raise |
| 116 | + except Exception as e: |
| 117 | + logger.error(f"[NVIDIA Embedding] Error: {e}", exc_info=True) |
| 118 | + raise |
| 119 | + |
| 120 | + def get_dim(self) -> int: |
| 121 | + if "embedding_dimensions" in self.provider_config: |
| 122 | + try: |
| 123 | + return int(self.provider_config["embedding_dimensions"]) |
| 124 | + except (ValueError, TypeError): |
| 125 | + logger.warning( |
| 126 | + f"embedding_dimensions in embedding configs is not a valid integer: " |
| 127 | + f"'{self.provider_config['embedding_dimensions']}', ignored." |
| 128 | + ) |
| 129 | + return 0 |
| 130 | + |
| 131 | + async def terminate(self): |
| 132 | + if self.client and not self.client.closed: |
| 133 | + await self.client.close() |
| 134 | + self.client = None |
0 commit comments