📄 中文摘要
在对 Karpathy 的 autoresearch 进行探索时，进行了两轮实验，共 60 次迭代。第一轮实验取得了改进，而第二轮实验则未能发现任何新成果，这一结果反而更具吸引力。实验使用了一个混合搜索系统，结合了 Cohere 嵌入和 pgvector 进行语义相似性计算，并在其上添加了关键词重排序层，使用 Django、PostgreSQL 和 Bedrock 技术栈。这种搜索架构在许多团队中广泛应用，实验结果为理解和优化搜索算法提供了重要见解。
我对一个生产搜索算法进行了 60 次 Autoresearch 实验。结果如何。
📄 English Summary
I Ran 60 Autoresearch Experiments on a Production Search Algorithm. Here's What Actually Happened.
Two rounds of experiments were conducted on Karpathy's autoresearch, totaling 60 iterations. The first round yielded improvements, while the second round found nothing, which turned out to be even more intriguing. The experiments utilized a hybrid search system that combines Cohere embeddings with pgvector for semantic similarity, along with a keyword re-ranking layer. The technology stack involved Django, PostgreSQL, and Bedrock, which is representative of many teams' current search architectures. The results provide valuable insights into understanding and optimizing search algorithms.