Recall@k 对比脚本
Recall@k 对比脚本
def recall_at_k(queries, k=5):
hit = 0
for q, gold_ids in queries:
ids = [h["doc_id"] for h in search(embed(q), top_k=k)]
hit += any(g in ids for g in gold_ids)
return hit / len(queries)
for name in ["bge-base-zh-v1.5", "bge-large-zh-v1.5"]:
print(name, recall_at_k(queries)) # queries 来自真实问题标注评测集 100-300 条真实问题即可起步,覆盖高频与长尾;对比时固定切片与索引参数,只变模型。