Open Deep Research on Cloudron - Self-Hosted AI Agent

[LoudLemur]

@robi said:

Is this still relevant?

I had a look.
https://github.com/assafelovic/gpt-researcher

seeems to be the most active, strongest option at the moment. It has a docker-compose. includes explicit “Deep Research” recursive/tree-like mode. Multi-LLM, local docs support, reports with citations. Very active

Or this:
https://github.com/dzhng/deep-research

Firecrawl for scraping, easy OpenAI-compatible

[robi]

@LoudLemur nice thanks.. a long while ago I was working on getting deepcode to work (until life went upside down), but now this seems more like a pipeline of projects that can be used to mashup various papers into a novel combination of pieces for use cases greater than the sum of its parts. Then back to generating a paper on the new discoveries.

[LoudLemur]

@robi Automated or ai-augmented research is very promising. If Macrohard goes well, we might soon have automated experimentation too, with the robots conducting experiments in labs. It is so exciting!

[timconsidine]

@loudlemur yes, an interesting area indeed.

I have a first package of dzhng/deepresearch for Cloudron.
Some opinionated implementations for Cloudron.
Testing.

[robi]

@LoudLemur Lol, is that the name of your company?

[LoudLemur]

@robi said:

@LoudLemur Lol, is that the name of your company?

I wish!

It is actually Elon Musk's company. He thought it was a funny pun on MicroSoft. It is a real company making real robot/ai hardware. He has been trying some automation with next gen vending machines and provisioning them.

https://x.com/elonmusk/status/1967800687431127265

https://x.com/MarioNawfal/status/1947318810597347654

Whilst we are on the topic, there is also a fun benchmark for how well a single AI can run an entire company:

https://ceobench.com/

[LoudLemur]

ooh, it was actually vending bench I had in mind for running a company:

https://andonlabs.com/evals/vending-bench-2

And there is another benchmark for the fun business of running a foodtruck in Austin:

https://foodtruckbench.com/

[robi]

@LoudLemur Yes vending bench is interesting.. several people are already running AI led companies, like Brian Roemmelle's Zero Human Company run by Grok as CEO, and others using CEO/Chief of staff/Manager level system prompts.

Nice to see which models are better at various subtasks, like Gemini Pro being a good negotiator.

Even free or local models might be decent at this if they were given better subtask skills and different system prompts for different roles.

Running larger than what fits in (v)RAM models on CPU only is also progressing as the different layers of the LLM are being separated from what needs to be in RAM vs what can be on SSD and loaded as needed.

[timconsidine]

I initially packaged https://github.com/dzhng/deep-research
But I had to add a front-end and I wasn't overly happy with it.
So trashed it and packaged https://github.com/u14app/deep-research which is a better implementation of the former (latter credits the former).

u14app version is smoother and works better.
Initially it defaulted to report writing in Chinese even when the request was in English (!). Known upstream issue. Hacked the package to force it respect system setting. But very sweet, it has a translation option, AND you can just ask it to rewrite the report in English (or whatever).

Now in Cloudron community app store listing
https://ca.cloudron.io/app/deep-research-u14
-or- :
https://communityapps.appx.uk/cloudron-deep-research/CloudronVersions.json

[robi]

@timconsidine Amazing, well done and thank you. Is there a dark mode too?

[timconsidine]

@robi yes, it is in the Settings (top right gear icon) under System : Theme is light, dark or auto.

I've got a few good reports out of it.
It even deals with something like :
Watch this video and make notes and a summary :

[robi]

@timconsidine great, thank you!

Would it benefit from something like this?
https://github.com/stretchcloud/deepscrape

[timconsidine]

@robi errr, conceptually yes but it seems to handle by itself. Can’t check how at the moment

[robi]

@timconsidine it may be more useful as a separate package connected to a local CPU only LLM to handle the non-linear needs.

I have a way to deploy tiny ones as a single file if need be.