10-engine system learning from each run; LLM ‘orchestrator’; open source APIs for 2,000+ vetted resources; AI-driven build-while-learning approach—enhanced with Google’s GenAI Processors architecture
How does this compare to OpenAI, Grok, and Perplexity "Deep Research" modes? You pulled 20x-40x more sources--did that translate to results that were 20x-40x better? Did you need to crawl that much data? How long did it take? Token count? If this targets scientists, do we want the system to run once, or should it do shorter, more focused searches with a human steering the process, like deep research usually works?
Deep Researches get 100 results combined, while this tool iterates and gets more results every time, going up to several thousands. The results’ quality is also evaluated and rated.
The saving is in terms of time. It produces a very detailed CSV.
This is applied findings for Alpha School researchers and curriculum scouts, for instance.
Also, what if we take https://github.com/assafelovic/gpt-researcher and tune some parameters up so that it does a broader search? Would it be roughly equivalent to what you system is doing?
Or run 10 gpt-researchers each configured to use a different model? And then combine the results?
What is the exact value of doing multi-model processing compared to using a single top-notch model?
Time: around 30 minutes, each comprehensive education run
Value: it uses Gemini's latest features and Firecrawl's multitude of tools to find 10,000s of links, which OpenAI, Anthropic, and Gemini engines evaluate by metadata and, if relevant and needed, explore the pdfs in the links to ensure the result is relevant. All wrapped in Gemini-Research Agent and careful python scripting. Everything is noted down, all the intermediate data is saved. Lastly, everything is collated and curated by an AI orchestrator (AI-in-the-loop), doing the jobs previously by a human, like planning, breaking down the task, and collating the report.
The value of the multi-modal specifically is that they are used both supplementary (Gemini Batch Crawl + Claude structured verifications, for instance)
and complimentary (Research-Agent (published less than a week ago and already incorporated here) + extensive Firecrawl integrations + OpenAI's latest and greatest of Web Search API).
This tool delivers more than just any GPT researcher by utilising newest features of multiple LLMs and my favorite part - the AI rchastrator makes sure you get high-quality results.
The choice is either to keep running limited-allowance Deep Search (watch out for duplication) or run this tool for 30 min, get 1000s of relevant results evaluated, rated, and collated into a csv.
Thanks for the implementation breakdown. You're throwing everything at it -- boiling, roasting, poaching, the whole kitchen sink. But how does the end result stack up next to something simple like a burger and fries or a basic pepperoni slice?
Would be helpful to see how this tool's output compares to other options. I get that might be extra lift, but unless the results can't be pulled off with other tools, the implementation details only go so far.
Link to the result by request
How does this compare to OpenAI, Grok, and Perplexity "Deep Research" modes? You pulled 20x-40x more sources--did that translate to results that were 20x-40x better? Did you need to crawl that much data? How long did it take? Token count? If this targets scientists, do we want the system to run once, or should it do shorter, more focused searches with a human steering the process, like deep research usually works?
Deep Researches get 100 results combined, while this tool iterates and gets more results every time, going up to several thousands. The results’ quality is also evaluated and rated.
The saving is in terms of time. It produces a very detailed CSV.
This is applied findings for Alpha School researchers and curriculum scouts, for instance.
Also, what if we take https://github.com/assafelovic/gpt-researcher and tune some parameters up so that it does a broader search? Would it be roughly equivalent to what you system is doing?
Or run 10 gpt-researchers each configured to use a different model? And then combine the results?
What is the exact value of doing multi-model processing compared to using a single top-notch model?
Good questions.
Time: around 30 minutes, each comprehensive education run
Value: it uses Gemini's latest features and Firecrawl's multitude of tools to find 10,000s of links, which OpenAI, Anthropic, and Gemini engines evaluate by metadata and, if relevant and needed, explore the pdfs in the links to ensure the result is relevant. All wrapped in Gemini-Research Agent and careful python scripting. Everything is noted down, all the intermediate data is saved. Lastly, everything is collated and curated by an AI orchestrator (AI-in-the-loop), doing the jobs previously by a human, like planning, breaking down the task, and collating the report.
The value of the multi-modal specifically is that they are used both supplementary (Gemini Batch Crawl + Claude structured verifications, for instance)
and complimentary (Research-Agent (published less than a week ago and already incorporated here) + extensive Firecrawl integrations + OpenAI's latest and greatest of Web Search API).
This tool delivers more than just any GPT researcher by utilising newest features of multiple LLMs and my favorite part - the AI rchastrator makes sure you get high-quality results.
The choice is either to keep running limited-allowance Deep Search (watch out for duplication) or run this tool for 30 min, get 1000s of relevant results evaluated, rated, and collated into a csv.
Thanks for the implementation breakdown. You're throwing everything at it -- boiling, roasting, poaching, the whole kitchen sink. But how does the end result stack up next to something simple like a burger and fries or a basic pepperoni slice?
Would be helpful to see how this tool's output compares to other options. I get that might be extra lift, but unless the results can't be pulled off with other tools, the implementation details only go so far.
Yes - I stated the same - your system handles more results and this seems to be the main difference. So how long does that actually take?
Example: my recent deep research with OpenAI finished in 8 minutes -- 27 sources, 108 searches.
How much time and money does your system need to handle 1,000 sources?
Did the Alpha School researchers test OpenAI deep research for the same task?
I'm trying to pin down clear business value here.