Embedding模型2026排行：中文检索场景实测

# 主指标：NDCG@10 (归一化折损累积增益)
# 辅助指标：Recall@5, Recall@10, MRR@10

def evaluate_retrieval(qrels, results, k=10):
    """计算检索评估指标"""
    ndcg_scores = []
    recall_scores = []
    mrr_scores = []
    
    for query_id, retrieved_docs in results.items():
        relevant_docs = qrels.get(query_id, {})
        
        # NDCG@k
        ndcg = compute_ndcg(retrieved_docs[:k], relevant_docs)
        ndcg_scores.append(ndcg)
        
        # Recall@k
        recall = len(set(retrieved_docs[:k]) & set(relevant_docs)) / len(relevant_docs)
        recall_scores.append(recall)
        
        # MRR@k
        mrr = compute_mrr(retrieved_docs[:k], relevant_docs)
        mrr_scores.append(mrr)
    
    return {
        "ndcg@10": np.mean(ndcg_scores),
        "recall@5": np.mean([compute_recall(r, q, 5) for r, q in zip(results, qrels)]),
        "recall@10": np.mean(recall_scores),
        "mrr@10": np.mean(mrr_scores),
    }

# 长文本检索对比
# 查询："解释GraphRAG的社区检测算法如何工作，以及它与传统RAG的区别"

# 检索到的文档长度分布
GTE-Qwen2-7B:  平均长度 2450字，命中率 92%
BGE-large-zh: 平均长度 850字，命中率 71%

# 结论：大模型Embedding能更好地理解长文档的全局语义

from FlagEmbedding import BGEM3FlagModel

model = BGEM3FlagModel(
    "BAAI/bge-m3",
    use_fp16=True,
    device="cuda"
)

# 支持三种检索方式
embeddings = model.encode(
    ["长文档内容..."],
    return_dense=True,      # 稠密向量
    return_sparse=True,     # 稀疏向量（BM25风格）
    return_colbert_vecs=True  # ColBERT向量（细粒度）
)

# 混合检索效果最佳
dense_score = cosine_similarity(query_emb, embeddings["dense"])
sparse_score = jaccard_similarity(query_sparse, embeddings["sparse"])
final_score = 0.7 * dense_score + 0.3 * sparse_score

from openai import OpenAI

client = OpenAI(api_key="your-key")

def get_embedding(text):
    response = client.embeddings.create(
        model="text-embedding-3-large",
        input=text,
        dimensions=3072,
    )
    return response.data[0].embedding

from FlagEmbedding import BGEM3FlagModel

model = BGEM3FlagModel("BAAI/bge-m3", use_fp16=True, device="cuda")

# 单句
embedding = model.encode(["你的文本"])["dense_vecs"][0]

# 批量（速度快）
embeddings = model.encode(
    ["文本1", "文本2", "文本3"],
    batch_size=12,
    max_length=8192,
)["dense_vecs"]

from transformers import AutoModel, AutoTokenizer
import torch

model = AutoModel.from_pretrained(
    "Alibaba-NLP/gte-Qwen2-7B-instruct",
    torch_dtype=torch.float16,
    trust_remote_code=True,
).cuda()

tokenizer = AutoTokenizer.from_pretrained(
    "Alibaba-NLP/gte-Qwen2-7B-instruct",
    trust_remote_code=True,
)

def get_embedding(text):
    # GTE-Qwen2支持指令式Embedding
    task = "为这个文档生成一个用于检索的向量表示"
    input_text = f"Instruct: {task}\nQuery: {text}"
    
    inputs = tokenizer(
        input_text, 
        return_tensors="pt", 
        max_length=32768,
        truncation=True,
    ).to("cuda")
    
    with torch.no_grad():
        embedding = model(**inputs).last_hidden_state[:, 0]
    
    return embedding[0].cpu().numpy()

import requests

API_URL = "https://api.jina.ai/v1/embeddings"
API_KEY = "your-key"

def get_embedding(text, task="retrieval.passage"):
    headers = {
        "Authorization": f"Bearer {API_KEY}",
        "Content-Type": "application/json",
    }
    
    data = {
        "model": "jina-embeddings-v3",
        "input": [text],
        "task": task,  # retrieval.query / retrieval.passage / separation
    }
    
    response = requests.post(API_URL, headers=headers, json=data)
    return response.json()["data"][0]["embedding"]

from sentence_transformers import SentenceTransformer, losses
from torch.utils.data import DataLoader

# 加载基座模型
model = SentenceTransformer("BAAI/bge-large-zh-v1.5")

# 准备训练数据：相似文档对
train_data = [
    {"query": "什么是RAG？", "positive": "RAG是检索增强生成...", "negative": "今天天气不错"},
    # ... 更多数据
]

# 损失函数
train_loss = losses.MultipleNegativesRankingLoss(model)

# 训练
model.fit(
    train_objectives=[(DataLoader(train_data), train_loss)],
    epochs=10,
    warmup_steps=100,
    output_path="./fine-tuned-embedding",
)