[vineyardmike]

I’m curious how this compares to the open-source version made by HuggingFace [1]. As I can tell, the HF version uses reasoning LLMs to search/traverse and parse the web and gather results, then evaluates the results before eventually synthesizing a result.

This version appears to show off a vector store for documents generated from a web crawl (the writer is a vector-store-aaS company)

[1] https://github.com/huggingface/smolagents/tree/main/examples...

[stefanwebb]

There's quite a few differences between HuggingFace's Open Deep-Research and Zilliz's DeepSearcher.

I think the biggest one is the goal: HF is to replicate the performance of Deep Research on the GAIA benchmark whereas ours is to teach agentic concepts and show how to build research agents with open-source.

Also, we go into the design in a lot more detail than HF's blog post. On the design side, HF uses code writing and execution as a tool, whereas we use prompt writing and calling as a tool. We do an explicit break down of the query into sub-queries, and sub-sub-queries, etc. whereas HF uses a chain of reasoning to decide what to do next.

I think ours is a better approach for producing a detailed report on an open-ended question, whereas HFs is better for answering a specific, challenging question in short form.