\n
pip install google-genai<\/code><\/pre>\n<\/div>\n\u30b9\u30c6\u30c3\u30d73\uff1a\u74b0\u5883\u5909\u6570\u306e\u8a2d\u5b9a<\/b><\/p>\n\n
export GEMINI_API_KEY=\"your-key-here\"<\/code><\/pre>\n<\/div>\n\u30b9\u30c6\u30c3\u30d74\uff1a\u30af\u30e9\u30a4\u30a2\u30f3\u30c8\u306e\u521d\u671f\u5316<\/b><\/code><\/p>\n<\/code><\/pre>\n<\/code><\/pre>\n\n
import os\nimport numpy as np \nfrom google import genai \nfrom google.genai import types \n\nclient = genai.Client(api_key=os.environ.get(\"GEMINI_API_KEY\")) \nMODEL = \"gemini-embedding-2-preview\"<\/span><\/code><\/pre>\n<\/div>\n<\/p>\n
\u30c6\u30ad\u30b9\u30c8\u57cb\u3081\u8fbc\u307f\u3068\u610f\u5473\u691c\u7d22<\/h2>\n
\u6700\u521d\u306e\u30b9\u30c6\u30c3\u30d7\u3068\u3057\u3066\u3001\u30c6\u30ad\u30b9\u30c8\u3060\u3051\u3067\u57fa\u672c\u7684\u306a\u610f\u5473\u691c\u7d22\u3092\u52d5\u304b\u3057\u3066\u307f\u307e\u3059\u3002<\/p>\n
\u30b9\u30c6\u30c3\u30d71\uff1a\u57cb\u3081\u8fbc\u307f\u95a2\u6570\u3068\u30b3\u30b5\u30a4\u30f3\u985e\u4f3c\u5ea6\u3092\u5b9a\u7fa9<\/b><\/p>\n<\/code><\/pre>\n\n
def embed_text(text: str, task_type: str = \"RETRIEVAL_DOCUMENT\", dims: int = 768): \n\u3000\u3000\"\"\"\u30c6\u30ad\u30b9\u30c8\u3092\u57cb\u3081\u8fbc\u307f\u30d9\u30af\u30c8\u30eb\u306b\u5909\u63db\"\"\" \n\u3000\u3000result = client.models.embed_content(\n model=MODEL,\n contents=text,\n config=types.EmbedContentConfig(\n task_type=task_type,\n output_dimensionality=dims,\n ),\n ) \n return result.embeddings[0].values \n\n\ndef cosine_sim(a, b):\n \"\"\"2\u3064\u306e\u30d9\u30af\u30c8\u30eb\u306e\u30b3\u30b5\u30a4\u30f3\u985e\u4f3c\u5ea6\u3092\u8a08\u7b97\uff08-1\u301c1\u30011\u306b\u8fd1\u3044\u307b\u3069\u985e\u4f3c\uff09\"\"\"\n a, b = np.array(a), np.array(b)\n return np.dot(a, b) \/ (np.linalg.norm(a) * np.linalg.norm(b))<\/code><\/pre>\n<\/div>\n<\/code><\/pre>\n\u30b9\u30c6\u30c3\u30d7\uff12\uff1a\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u3092\u30a4\u30f3\u30c7\u30c3\u30af\u30b9\u5316<\/b><\/p>\n<\/code><\/pre>\n\n
docs = [\n \"\u6771\u4eac\u30bf\u30ef\u30fc\u306f1958\u5e74\u306b\u5b8c\u6210\u3057\u305f\u8d64\u3044\u9244\u5854\u3067\u3059\u3002\",\n \"\u30b9\u30ab\u30a4\u30c4\u30ea\u30fc\u306f634\u30e1\u30fc\u30c8\u30eb\u306e\u4e16\u754c\u4e00\u9ad8\u3044\u96fb\u6ce2\u5854\u3067\u3059\u3002\",\n \"\u5bff\u53f8\u306f\u65e5\u672c\u3092\u4ee3\u8868\u3059\u308b\u6599\u7406\u306e\u3072\u3068\u3064\u3067\u3059\u3002\",\n \"\u5bcc\u58eb\u5c71\u306f\u6a19\u9ad83776\u30e1\u30fc\u30c8\u30eb\u306e\u65e5\u672c\u6700\u9ad8\u5cf0\u3067\u3059\u3002\",\n]\ndoc_vecs = [embed_text(d, task_type=\"RETRIEVAL_DOCUMENT\") for d in docs]\n<\/code><\/pre>\n<\/div>\n<\/code><\/pre>\n\u30b9\u30c6\u30c3\u30d73\uff1a\u30c6\u30ad\u30b9\u30c8\u30af\u30a8\u30ea\u3067\u691c\u7d22<\/b><\/p>\n<\/code><\/pre>\n\n
query = \"\u6771\u4eac\u306e\u9ad8\u3044\u5efa\u7269\u306b\u3064\u3044\u3066\u6559\u3048\u3066\" \nq_vec = embed_text(query, task_type=\"RETRIEVAL_QUERY\") \n\nscores = [(cosine_sim(q_vec, dv), doc) for dv, doc in zip(doc_vecs, docs)] \nfor score, doc in sorted(scores, reverse=True): \n print(f\" {score:.4f} | {doc}\")<\/code><\/pre>\n<\/div>\n<\/code><\/pre>\n\u30b9\u30c6\u30c3\u30d7\uff14\uff1a\u5b9f\u884c\u7d50\u679c<\/b><\/p>\n
\u30af\u30a8\u30ea\u300c\u6771\u4eac\u306e\u9ad8\u3044\u5efa\u7269\u300d\u306e\u691c\u7d22\u7d50\u679c\u306f\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u306a\u308a\u307e\u3057\u305f\u3002<\/p>\n
<\/code><\/pre>\n\n
\u30af\u30a8\u30ea: \u300c\u6771\u4eac\u306e\u9ad8\u3044\u5efa\u7269\u300d\n\n--- \u610f\u5473\u691c\u7d22\u7d50\u679c ---\n 0.7180 | \u6771\u4eac\u30bf\u30ef\u30fc\u306f1958\u5e74\u306b\u5b8c\u6210\u3057\u305f\u8d64\u3044\u9244\u5854\u3067\u3059\u3002\n 0.7176 | \u30b9\u30ab\u30a4\u30c4\u30ea\u30fc\u306f634\u30e1\u30fc\u30c8\u30eb\u306e\u4e16\u754c\u4e00\u9ad8\u3044\u96fb\u6ce2\u5854\u3067\u3059\u3002\n 0.6541 | \u5bcc\u58eb\u5c71\u306f\u6a19\u9ad83776\u30e1\u30fc\u30c8\u30eb\u306e\u65e5\u672c\u6700\u9ad8\u5cf0\u3067\u3059\u3002\n 0.5491 | \u5bff\u53f8\u306f\u65e5\u672c\u3092\u4ee3\u8868\u3059\u308b\u6599\u7406\u306e\u3072\u3068\u3064\u3067\u3059\u3002\n<\/code><\/pre>\n<\/div>\n<\/code><\/pre>\n\u300c\u6771\u4eac\u306e\u9ad8\u3044\u5efa\u7269\u300d\u3068\u3044\u3046\u30af\u30a8\u30ea\u306b\u5bfe\u3057\u3066\u3001\u300c\u30b9\u30ab\u30a4\u30c4\u30ea\u30fc\u300d\u300c\u6771\u4eac\u30bf\u30ef\u30fc\u300d\u304c\u4e0a\u4f4d\u306b\u6765\u3066\u3044\u307e\u3059\u3002\u30af\u30a8\u30ea\u306b\u300c\u30b9\u30ab\u30a4\u30c4\u30ea\u30fc\u300d\u3068\u3044\u3046\u5358\u8a9e\u304c\u542b\u307e\u308c\u3066\u3044\u306a\u304f\u3066\u3082\u610f\u5473\u306e\u8fd1\u3055\u3067\u30d2\u30c3\u30c8\u3059\u308b\u3002\u3053\u308c\u304c\u30ad\u30fc\u30ef\u30fc\u30c9\u691c\u7d22\u3068\u306e\u6c7a\u5b9a\u7684\u306a\u9055\u3044\u3067\u3059\u3002\u300c\u5bcc\u58eb\u5c71\u300d\u3082\u300c\u9ad8\u3044\u300d\u3068\u3044\u3046\u610f\u5473\u7684\u306a\u7e4b\u304c\u308a\u3067\u4e2d\u7a0b\u5ea6\u306e\u30b9\u30b3\u30a2\u304c\u4ed8\u304d\u3001\u7121\u95a2\u4fc2\u306a\u300c\u5bff\u53f8\u300d\u306f\u3057\u3063\u304b\u308a\u4f4e\u30b9\u30b3\u30a2\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n
<\/p>\n
\u753b\u50cf\u306e\u57cb\u3081\u8fbc\u307f\u3068\u30af\u30ed\u30b9\u30e2\u30fc\u30c0\u30eb\u691c\u7d22<\/h2>\n
\u3053\u3053\u304b\u3089\u304cGemini Embedding 2\u306e\u771f\u9aa8\u9802\u3067\u3059\u3002\u753b\u50cf\u3082\u30c6\u30ad\u30b9\u30c8\u3068\u540c\u3058\u30d9\u30af\u30c8\u30eb\u7a7a\u9593\u306b\u57cb\u3081\u8fbc\u3081\u308b\u305f\u3081\u3001\u30c6\u30ad\u30b9\u30c8\u3067\u753b\u50cf\u3092\u691c\u7d22\u3067\u304d\u307e\u3059\u3002<\/span><\/p>\n\u30b9\u30c6\u30c3\u30d7\uff11\uff1a\u753b\u50cf\u57cb\u3081\u8fbc\u307f\u95a2\u6570\u3092\u5b9a\u7fa9<\/b><\/p>\n<\/code><\/pre>\n\n
from pathlib import Path\n\ndef embed_image(image_path: str) -> list[float]:\n \"\"\"1\u679a\u306e\u753b\u50cf \u2192 \u30d9\u30af\u30c8\u30eb\"\"\"\n with open(image_path, \"rb\") as f:\n data = f.read()\n\n suffix = Path(image_path).suffix.lower()\n mime = \"image\/jpeg\" if suffix in (\".jpg\", \".jpeg\") else \"image\/png\"\n\n result = client.models.embed_content(\n model=MODEL,\n contents=[types.Part.from_bytes(data=data, mime_type=mime)],\n config=types.EmbedContentConfig(\n task_type=\"RETRIEVAL_DOCUMENT\",\n output_dimensionality=768,\n ),\n )\n return result.embeddings[0].values\n<\/code><\/pre>\n<\/div>\n<\/code><\/pre>\n\u753b\u50cf\u306e\u6e21\u3057\u65b9\u306f\u30b7\u30f3\u30d7\u30eb\u3067\u3001\u30d5\u30a1\u30a4\u30eb\u3092\u30d0\u30a4\u30c8\u5217\u3067\u8aad\u307f\u8fbc\u3093\u3067 Part.from_bytes<\/code> \u306b\u6e21\u3059\u3060\u3051\u3067\u3059\u3002PIL.Image\u3078\u306e\u5909\u63db\u306f\u4e0d\u8981\u3067\u3059\u3002<\/p>\n\u30b9\u30c6\u30c3\u30d72\uff1a\u30d5\u30a9\u30eb\u30c0\u5185\u306e\u753b\u50cf\u3092\u30a4\u30f3\u30c7\u30c3\u30af\u30b9\u5316<\/b><\/p>\n<\/code><\/pre>\n\n
def build_image_index(image_dir: str) -> list[dict]: \n\u3000 \"\"\"\u30d5\u30a9\u30eb\u30c0\u5185\u306e\u753b\u50cf\u3092\u5168\u3066\u30d9\u30af\u30c8\u30eb\u5316\"\"\" \n index = [] \n extensions = {\".jpg\", \".jpeg\", \".png\"} \n paths = [p for p in Path(image_dir).iterdir() if p.suffix.lower() in extensions] \n\n for path in paths: \n vec = embed_image(str(path)) \n index.append({\"path\": str(path), \"vector\": vec}) \n print(f\" indexed: {path.name}\") \n return index<\/code><\/pre>\n<\/div>\n<\/code><\/pre>\n\u30b9\u30c6\u30c3\u30d73\uff1a\u30c6\u30ad\u30b9\u30c8\u30af\u30a8\u30ea\u3067\u753b\u50cf\u3092\u691c\u7d22<\/b><\/p>\n<\/code><\/pre>\n\n
def search_images_by_text(query: str, index: list[dict], top_k: int = 5): \n \"\"\"\u30c6\u30ad\u30b9\u30c8\u30af\u30a8\u30ea\u3067\u753b\u50cf\u3092\u691c\u7d22\"\"\" \n result = client.models.embed_content( \n model=MODEL, \n contents=query, \n config=types.EmbedContentConfig( \n task_type=\"RETRIEVAL_QUERY\", \n output_dimensionality=768, \n ), \n ) \n q_vec = result.embeddings[0].values \n\n scored = [] \n for item in index: \n score = cosine_sim(q_vec, item[\"vector\"]) \n scored.append((score, item[\"path\"])) \n\n scored.sort(reverse=True) \n return scored[:top_k]<\/code><\/pre>\n<\/div>\n<\/code><\/pre>\n\u30b9\u30c6\u30c3\u30d74\uff1a\u5b9f\u884c\u7d50\u679c<\/b><\/p>\n
7\u679a\u306e\u753b\u50cf\uff08\u6771\u4eac\u30bf\u30ef\u30fc\u00d72\u3001\u30b9\u30ab\u30a4\u30c4\u30ea\u30fc\u3001\u5bcc\u58eb\u5c71\u3001\u5bff\u53f8\u3001\u308a\u3093\u3054\u3001\u732b\uff09\u3092\u7528\u610f\u3057\u3066\u30c6\u30ad\u30b9\u30c8\u30af\u30a8\u30ea\u3067\u691c\u7d22\u3057\u307e\u3057\u305f\u3002<\/p>\n
\n- \u300c\u98df\u3079\u7269\u306e\u5199\u771f\u300d<\/strong> \u3067\u691c\u7d22\uff1a<\/li>\n<\/ol>\n
<\/code><\/pre>\n\n
\u30af\u30a8\u30ea: \u300c\u98df\u3079\u7269\u306e\u5199\u771f\u300d\n 0.3817 | sushi.jpg\n 0.3395 | apple.jpg\n 0.3085 | Fuji.jpg 0.3007 | cat.jpg\n 0.2979 | tokyotower_black.jpg\n 0.2951 | tokyotower_blue.jpg\n 0.2653 | skytree.jpg<\/code><\/pre>\n<\/div>\n<\/code><\/pre>\n\u5bff\u53f8\u3068\u308a\u3093\u3054\u304c\u30c8\u30c3\u30d72\u3002\u98df\u3079\u7269\u306b\u95a2\u9023\u3059\u308b\u753b\u50cf\u304c\u4e0a\u4f4d\u306b\u6765\u3066\u3044\u307e\u3059\u3002<\/p>\n
2. \u300c\u8d64\u3044\u3082\u306e\u300d<\/strong> \u3067\u691c\u7d22\uff1a<\/p>\n<\/code><\/pre>\n\n
\u30af\u30a8\u30ea: \u300c\u8d64\u3044\u3082\u306e\u300d\n 0.3705 | apple.jpg\n 0.3474 | tokyotower_blue.jpg\n 0.3363 | tokyotower_black.jpg\n 0.3295 | sushi.jpg\n 0.3018 | Fuji.jpg\n 0.2864 | cat.jpg\n 0.2825 | skytree.jpg<\/span><\/code><\/pre>\n<\/div>\n<\/code><\/pre>\n\u8d64\u3044\u308a\u3093\u3054\u3001\u8d64\u3044\u6771\u4eac\u30bf\u30ef\u30fc\u304c\u4e0a\u4f4d\u3002\u8272\u3068\u3044\u3046\u62bd\u8c61\u7684\u306a\u6982\u5ff5\u3067\u3082\u753b\u50cf\u3092\u5f15\u3063\u5f35\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n
3. \u300c\u304b\u308f\u3044\u3044\u3082\u306e\u300d<\/strong> \u3067\u691c\u7d22\uff1a<\/p>\n<\/code><\/pre>\n\n
\u30af\u30a8\u30ea: \u300c\u304b\u308f\u3044\u3044\u3082\u306e\u300d\n 0.3273 | cat.jpg\n 0.3183 | sushi.jpg\n 0.2986 | apple.jpg\n 0.2956 | Fuji.jpg\n 0.2859 | tokyotower_black.jpg\n 0.2844 | tokyotower_blue.jpg\n 0.2716 | skytree.jpg<\/code><\/pre>\n<\/div>\n<\/code><\/pre>\n\u732b\u304c\u30c8\u30c3\u30d7\u3002\u300c\u304b\u308f\u3044\u3044\u300d\u306e\u3088\u3046\u306a\u4e3b\u89b3\u7684\u30fb\u611f\u899a\u7684\u306a\u30af\u30a8\u30ea\u3067\u3082\u3001\u305d\u308c\u306a\u308a\u306b\u610f\u5473\u306e\u3042\u308b\u7d50\u679c\u304c\u8fd4\u3063\u3066\u304f\u308b\u306e\u306f\u9762\u767d\u3044\u3068\u3053\u308d\u3067\u3059\u3002<\/p>\n
<\/p>\n
\u30c6\u30ad\u30b9\u30c8\uff0b\u753b\u50cf\u306e\u8907\u5408\u57cb\u3081\u8fbc\u307f<\/h2>\n
Gemini Embedding 2\u306f\u3001\u30c6\u30ad\u30b9\u30c8\u3068\u753b\u50cf\u3092\u7d44\u307f\u5408\u308f\u305b\u30661\u3064\u306e\u7d71\u5408\u30d9\u30af\u30c8\u30eb\u3092\u751f\u6210\u3059\u308b\u3053\u3068\u3082\u3067\u304d\u307e\u3059\u3002\u753b\u50cf\u3060\u3051\u306e\u57cb\u3081\u8fbc\u307f\u3068\u6bd4\u3079\u3066\u3069\u308c\u304f\u3089\u3044\u7cbe\u5ea6\u304c\u5909\u308f\u308b\u306e\u304b\u3001\u5b9f\u969b\u306b\u6bd4\u8f03\u3057\u3066\u307f\u307e\u3057\u305f\u3002<\/p>\n
\u30b9\u30c6\u30c3\u30d71\uff1a\u8907\u5408\u57cb\u3081\u8fbc\u307f\u95a2\u6570\u3092\u5b9a\u7fa9<\/b><\/p>\n<\/code><\/pre>\n\n
def embed_text_and_image(text: str, image_path: str) -> list[float]:\n \"\"\"\u30c6\u30ad\u30b9\u30c8 + \u753b\u50cf \u2192 1\u3064\u306e\u7d71\u5408\u30d9\u30af\u30c8\u30eb\"\"\"\n with open(image_path, \"rb\") as f:\n data = f.read()\n\n suffix = Path(image_path).suffix.lower()\n mime = \"image\/jpeg\" if suffix in (\".jpg\", \".jpeg\") else \"image\/png\"\n\n result = client.models.embed_content(\n model=MODEL,\n contents=[\n types.Content(\n parts=[\n types.Part(text=text),\n types.Part.from_bytes(data=data, mime_type=mime),\n ]\n )\n ],\n config=types.EmbedContentConfig(\n output_dimensionality=768,\n ),\n )\n return result.embeddings[0].values\n<\/code><\/pre>\n<\/div>\n<\/code><\/pre>\n\u4eca\u56de\u306f\u30d5\u30a1\u30a4\u30eb\u540d\u3092\u305d\u306e\u307e\u307e\u30c6\u30ad\u30b9\u30c8\u3068\u3057\u3066\u5165\u529b\u3057\u3001\u753b\u50cf\u306e\u307f\u306e\u57cb\u3081\u8fbc\u307f\u3068\u6bd4\u8f03\u3057\u307e\u3057\u305f\u3002<\/p>\n
<\/b>
\u30b9\u30c6\u30c3\u30d72\uff1a\u5b9f\u884c\u7d50\u679c\uff08\u753b\u50cf\u306e\u307f vs \u30c6\u30ad\u30b9\u30c8\uff0b\u753b\u50cf\uff09<\/b><\/p>\n\n- \u300c\u98df\u3079\u7269\u306e\u5199\u771f\u300d\u3067\u691c\u7d22<\/li>\n<\/ol>\n
\u30af\u30a8\u30ea: \u300c\u98df\u3079\u7269\u306e\u5199\u771f\u300d\n [\u753b\u50cf\u306e\u307f] [\u30c6\u30ad\u30b9\u30c8\uff0b\u753b\u50cf]\n 0.3817 sushi.jpg 0.4735 sushi.jpg\n 0.3395 query.jpg 0.4261 apple.jpg\n 0.3395 apple.jpg 0.4216 query.jpg\n 0.3085 Fuji.jpg 0.3808 cat.jpg\n 0.3007 cat.jpg 0.3762 Fuji.jpg\n 0.2979 tokyotower_black.jpg 0.3504 skytree.jpg\n 0.2951 tokyotower_blue.jpg 0.3461 tokyotower_blue.jpg\n<\/code><\/pre>\n\n- \u300c\u8d64\u3044\u3082\u306e\u300d\u3067<\/strong>\u691c\u7d22<\/li>\n<\/ol>\n
\u30af\u30a8\u30ea: \u300c\u8d64\u3044\u3082\u306e\u300d\n [\u753b\u50cf\u306e\u307f] [\u30c6\u30ad\u30b9\u30c8\uff0b\u753b\u50cf]\n 0.3705 query.jpg 0.4484 apple.jpg\n 0.3705 apple.jpg 0.4417 query.jpg\n 0.3474 tokyotower_blue.jpg 0.3999 sushi.jpg\n 0.3363 tokyotower_black.jpg 0.3845 tokyotower_black.jpg\n 0.3295 sushi.jpg 0.3818 tokyotower_blue.jpg \n 0.3018 Fuji.jpg 0.3772 cat.jpg\n 0.2864 cat.jpg 0.3683 skytree.jpg\n<\/code><\/pre>\n\n- \u300c\u304b\u308f\u3044\u3044\u3082\u306e\u300d\u3067\u691c\u7d22<\/li>\n<\/ol>\n
\u30af\u30a8\u30ea: \u300c\u304b\u308f\u3044\u3044\u3082\u306e\u300d\n [\u753b\u50cf\u306e\u307f] [\u30c6\u30ad\u30b9\u30c8\uff0b\u753b\u50cf]\n 0.3273 cat.jpg 0.4007 cat.jpg\n 0.3183 sushi.jpg 0.3973 sushi.jpg\n 0.2986 query.jpg 0.3930 apple.jpg\n 0.2986 apple.jpg 0.3823 query.jpg\n 0.2956 Fuji.jpg 0.3683 skytree.jpg\n 0.2859 tokyotower_black.jpg 0.3449 tokyotower_black.jp\n 0.2844 tokyotower_blue.jpg 0.3411 Fuji.jpg\n<\/code><\/pre>\n\u5168\u4f53\u7684\u306b\u30c6\u30ad\u30b9\u30c8\uff0b\u753b\u50cf\u306e\u8907\u5408\u57cb\u3081\u8fbc\u307f\u306e\u307b\u3046\u304c\u30b9\u30b3\u30a2\u304c\u9ad8\u304f\u51fa\u3066\u304a\u308a\u3001\u6700\u5c0f\u9650\u306e\u30c6\u30ad\u30b9\u30c8\u60c5\u5831\u3092\u8db3\u3057\u305f\u3060\u3051\u3067\u3082\u3001\u30b9\u30b3\u30a2\u304c\u660e\u78ba\u306b\u5411\u4e0a\u3057\u307e\u3057\u305f\u3002<\/p>\n
<\/p>\n
<\/p>\n
\u5b9f\u969b\u306e\u30e6\u30fc\u30b9\u30b1\u30fc\u30b9<\/h2>\n
\u4eca\u56de\u8a66\u3057\u305f\u30c6\u30ad\u30b9\u30c8\u691c\u7d22\u30fb\u753b\u50cf\u691c\u7d22\u306e\u4ed6\u306b\u3082\u3001\u3055\u307e\u3056\u307e\u306a\u30e6\u30fc\u30b9\u30b1\u30fc\u30b9\u304c\u8003\u3048\u3089\u308c\u307e\u3059\u3002<\/p>\n
\n- \u30de\u30eb\u30c1\u30e2\u30fc\u30c0\u30ebRAG \uff1a\u793e\u5185\u30de\u30cb\u30e5\u30a2\u30eb\uff08PDF\uff09\u3068\u88fd\u54c1\u5199\u771f\u3092\u307e\u3068\u3081\u3066\u30a4\u30f3\u30c7\u30c3\u30af\u30b9\u3057\u3001\u300c\u8d64\u3044\u90e8\u54c1\u306e\u53d6\u308a\u4ed8\u3051\u65b9\u300d\u3067\u95a2\u9023\u3059\u308bPDF\u30da\u30fc\u30b8\u3068\u5199\u771f\u306e\u4e21\u65b9\u3092\u30d2\u30c3\u30c8\u3055\u305b\u308b\u6a2a\u65ad\u691c\u7d22\u304c\u53ef\u80fd\u306b\u306a\u308a\u307e\u3059\u3002<\/li>\n
- \u5546\u54c1\u753b\u50cf\u691c\u7d22\uff1a\u300c\u3075\u308f\u3075\u308f\u306e\u767d\u3044\u30bb\u30fc\u30bf\u30fc\u300d\u306e\u3088\u3046\u306a\u81ea\u7136\u8a00\u8a9e\u304b\u3089\u5546\u54c1\u753b\u50cf\u3092\u76f4\u63a5\u691c\u7d22\u3067\u304d\u307e\u3059\u3002\u5546\u54c1\u8aac\u660e\u30c6\u30ad\u30b9\u30c8\u3092\u4e8b\u524d\u306b\u7528\u610f\u3059\u308b\u624b\u9593\u304c\u7701\u3051\u307e\u3059\u3002<\/li>\n
- \u52d5\u753b\u30fb\u97f3\u58f0\u306e\u30bb\u30de\u30f3\u30c6\u30a3\u30c3\u30af\u691c\u7d22\uff1a\u4f1a\u8b70\u9332\u753b\u3084\u8b1b\u7fa9\u52d5\u753b\u3092\u30d9\u30af\u30c8\u30eb\u5316\u3057\u3066\u304a\u3051\u3070\u3001\u30c6\u30ad\u30b9\u30c8\u3067\u4e2d\u8eab\u3092\u691c\u7d22\u3067\u304d\u307e\u3059\u3002<\/li>\n<\/ol>\n
<\/p>\n
\u6ce8\u610f\u70b9<\/h2>\n
\u5c0e\u5165\u524d\u306b\u62bc\u3055\u3048\u3066\u304a\u304f\u3079\u304d\u5236\u7d04\u304c\u3044\u304f\u3064\u304b\u3042\u308a\u307e\u3059\u3002<\/p>\n
\n- \u30fb\u30ea\u30fc\u30b8\u30e7\u30f3\u5236\u9650\uff1aVertex AI\u3067\u306e\u63d0\u4f9b\u306f\u73fe\u6642\u70b9\u3067us-central1\u30ea\u30fc\u30b8\u30e7\u30f3\u306e\u307f\u3067\u3001\u65e5\u672c\u30ea\u30fc\u30b8\u30e7\u30f3\uff08asia-northeast1\uff09\u306f\u975e\u5bfe\u5fdc\u3067\u3059\u3002Gemini API\u7d4c\u7531\u3067\u3042\u308c\u3070\u30ea\u30fc\u30b8\u30e7\u30f3\u306e\u5236\u7d04\u306f\u3042\u308a\u307e\u305b\u3093\u3002<\/li>\n
- \u30fb\u65e2\u5b58\u30c7\u30fc\u30bf\u3068\u306e\u4e92\u63db\u6027\uff1agemini-embedding-001\u3068\u306f\u57cb\u3081\u8fbc\u307f\u7a7a\u9593\u306b\u4e92\u63db\u6027\u304c\u306a\u3044\u305f\u3081\u3001\u79fb\u884c\u6642\u306f\u5168\u30c7\u30fc\u30bf\u306e\u518d\u57cb\u3081\u8fbc\u307f\u304c\u5fc5\u8981\u3067\u3059\u3002<\/li>\n
- \u30fb\u30d7\u30ec\u30d3\u30e5\u30fc\u6bb5\u968e\uff1a\u73fe\u5728\u30d1\u30d6\u30ea\u30c3\u30af\u30d7\u30ec\u30d3\u30e5\u30fc\u306e\u305f\u3081\u3001API\u306e\u4ed5\u69d8\u304c\u5909\u66f4\u3055\u308c\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002\u672c\u756a\u74b0\u5883\u3078\u306e\u5c0e\u5165\u306f\u4e00\u822c\u63d0\u4f9b\u5f8c\u304c\u5b89\u5168\u3067\u3059\u3002<\/li>\n<\/ul>\n
<\/p>\n
\u307e\u3068\u3081<\/h2>\n
Gemini Embedding 2\u306f\u3001\u30c6\u30ad\u30b9\u30c8\u3068\u753b\u50cf\uff08\u3055\u3089\u306b\u306f\u52d5\u753b\u30fb\u97f3\u58f0\u30fbPDF\uff09\u3092\u540c\u3058\u30d9\u30af\u30c8\u30eb\u7a7a\u9593\u306b\u7d71\u5408\u3067\u304d\u308b\u3068\u3044\u3046\u70b9\u3067\u3001\u57cb\u3081\u8fbc\u307f\u30e2\u30c7\u30eb\u306e\u4f7f\u3044\u65b9\u3092\u5927\u304d\u304f\u5909\u3048\u308b\u30dd\u30c6\u30f3\u30b7\u30e3\u30eb\u3092\u6301\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n
\u5b9f\u969b\u306b\u8a66\u3057\u3066\u307f\u308b\u3068\u3001API\u306e\u4f7f\u3044\u52dd\u624b\u306f\u30b7\u30f3\u30d7\u30eb\u3067\u3001\u6570\u884c\u306e\u30b3\u30fc\u30c9\u3067\u30af\u30ed\u30b9\u30e2\u30fc\u30c0\u30eb\u691c\u7d22\u304c\u52d5\u304d\u307e\u3059\u3002\u7279\u306b\u300c\u30c6\u30ad\u30b9\u30c8\u3067\u753b\u50cf\u3092\u691c\u7d22\u3067\u304d\u308b\u300d\u4f53\u9a13\u306f\u3001\u5f93\u6765\u306eCLIP\u30d9\u30fc\u30b9\u306e\u30a2\u30d7\u30ed\u30fc\u30c1\u3068\u6bd4\u3079\u3066\u3082\u30bb\u30c3\u30c8\u30a2\u30c3\u30d7\u304c\u5727\u5012\u7684\u306b\u7c21\u5358\u3067\u3059\u3002<\/p>\n
\u307e\u305a\u306f\u624b\u5143\u306e\u753b\u50cf\u30d5\u30a9\u30eb\u30c0\u3067\u8a66\u3057\u3066\u307f\u3066\u3001\u30de\u30eb\u30c1\u30e2\u30fc\u30c0\u30ebRAG\u3084\u5546\u54c1\u691c\u7d22\u3078\u306e\u5fdc\u7528\u3092\u691c\u8a0e\u3057\u3066\u307f\u3066\u306f\u3044\u304b\u304c\u3067\u3057\u3087\u3046\u304b\u3002<\/p>\n
<\/p>\n
\u53c2\u8003<\/h2>\n
Google Cloud Documentation<\/a><\/p>","protected":false},"excerpt":{"rendered":"Google\u767a\u306e\u30de\u30eb\u30c1\u30e2\u30fc\u30c0\u30eb\u57cb\u3081\u8fbc\u307f\u30e2\u30c7\u30eb\u300cGemini Embedding 2\u300d\u3092\u5b9f\u969b\u306b\u8a66\u3057\u3066\u307f\u307e\u3057\u305f\u3002\u30c6\u30ad\u30b9\u30c8\u30fb\u753b\u50cf\u3092\u540c\u3058\u30d9\u30af\u30c8\u30eb\u7a7a\u9593\u306b\u57cb\u3081\u8fbc\u307f\u3001\u30c6\u30ad\u30b9\u30c8\u3067\u753b\u50cf\u3092\u691c\u7d22\u3059\u308b\u30af\u30ed\u30b9\u30e2\u30fc\u30c0\u30eb\u691c\u7d22\u3092Python\u30b3\u30fc\u30c9\u4ed8\u304d\u3067 […]<\/p>\n","protected":false},"author":29,"featured_media":7729,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"content-type":"","swell_btn_cv_data":"","footnotes":"","_wp_rev_ctl_limit":""},"categories":[1249],"tags":[96,103,331,39,26,419],"class_list":["post-7720","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-knowledge","tag-ai","tag-google","tag-python","tag-39","tag-26","tag-419"],"_links":{"self":[{"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/posts\/7720","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/users\/29"}],"replies":[{"embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/comments?post=7720"}],"version-history":[{"count":3,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/posts\/7720\/revisions"}],"predecessor-version":[{"id":8404,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/posts\/7720\/revisions\/8404"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/media\/7729"}],"wp:attachment":[{"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/media?parent=7720"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/categories?post=7720"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/tags?post=7720"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}