Hi all! Aymeric (m-ric) here, maintainer of smolagents and part of the team who built this. Happy to see this interesting people here!
Few points:
- open Deep Research is not a production app, but it could easily be productionized (would need to be faster + good UX).
- As the GAIA score of 55% (not 54%, that would be lame) says, it's not far from the Deep Research score of 67%. It's also not there yet: I think the main point of progress is to improve web browsing. We're working on integrating vision models (for now we've used a text browser developed by the Microsofit autogen team, congrats to them) because it's probably the best way to really interact with webpages.
- Open Deep Research is built on smolagents, a library that we're building, for which the core is having agents that write their actions (tool calls) in code snippets instead of the unpractical JSON blobs + parsing that everyone incl OpenAI and Anthropic use for their agentic/tool-calling APIs. Don't hesitate to go try out the lib and drop issues/PRs!
- smolagents does code execution, which means "danger for your machine" if ran locally. We've railguardeed that a bit with our custom python interpreter, but it will never be 100% safe, so we're enabling remote execution with E2B and soon Docker.
> smolagents does code execution, which means "danger for your machine" if ran locally. We've railguardeed that a bit with our custom python interpreter, but it will never be 100% safe, so we're enabling remote execution with E2B and soon Docker.
Those remote interfaces may also work with local VMs for isolation.
Yeah, that's what I was thinking: just throw the whole lot inside a Docker container and call it a day. Unless you're dealing with potentially malicious code that could break out of a container, that should isolate the rest of your machine sufficiently.
Alternatively, PyPy is actually fully sandboxable.
Great work on this, Aymeric and team! In terms of improving browsing and/or data sources, do you think it might be worth integrating things like Google Scholar search capability to increase the depth of some of the research that can be done?
It's something I'd be happy to explore a bit if it's of interest.
> for now we've used a text browser developed by the Microsofit autogen team, congrats to them
oh super cool! i've usually heard it the other way - people develop LLM-friendly web scrapers. i wrote one for myself, and for others there's firecrawl and expand.ai. a full "text browser" (i guess with rendering?) run locally seems like a better solution for local agents.
I think using vision models for browsing is the wrong approach. It is the same as using OCR for scanning PDFs. The underlying text is already in digital form. So it would make more sense to establish a standard similar to meta-tags that enable the agentic web.
If you're working from the markup rather than the appearance of the page, you're probably increasing the incentives for metacrap, "invisible text spam" and similar tactics.
PDFs are more akin to SVG than to a Word document, and the text is often very far from “available”. OCR can be the only way to reconstruct the document as it appears on screen.
Few points:
- open Deep Research is not a production app, but it could easily be productionized (would need to be faster + good UX).
- As the GAIA score of 55% (not 54%, that would be lame) says, it's not far from the Deep Research score of 67%. It's also not there yet: I think the main point of progress is to improve web browsing. We're working on integrating vision models (for now we've used a text browser developed by the Microsofit autogen team, congrats to them) because it's probably the best way to really interact with webpages.
- Open Deep Research is built on smolagents, a library that we're building, for which the core is having agents that write their actions (tool calls) in code snippets instead of the unpractical JSON blobs + parsing that everyone incl OpenAI and Anthropic use for their agentic/tool-calling APIs. Don't hesitate to go try out the lib and drop issues/PRs!
- smolagents does code execution, which means "danger for your machine" if ran locally. We've railguardeed that a bit with our custom python interpreter, but it will never be 100% safe, so we're enabling remote execution with E2B and soon Docker.