news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

The academic search engine created by a three-person team directly won the SOTA!

It claims that the relevance of search results is on average 5 times higher than that of Google Scholar, can search over 100 million research objects, and does not rely on LLM.

What's more important:free, the students are ecstatic!

Let’s take a look at the results. Enter the search keyword and it only takes 350 milliseconds for the highly relevant documents to be listed:

You can also filter out viewable PDF files with one click.

From year, number of citations, to publication type and specific journal, you can set one-click filtering:

Click the title of the searched paper to preview it, and it supports one-click copying of common academic citation formats:

This new search engine is calledLumina, it is said that it has processed more than 300,000 queries and supports 24 languages.

In order to compare with traditional academic search engines such as Google Scholar, the development team also conducted a special benchmark test, and the code base has been open sourced. The test results show that Lumina's search results ranked SOTA in terms of relevance.

The developer tweeted to challenge Google:

It took the three of us a few months to produce it, and it was better than Google's.

After watching the video, AI 3D simulation engine Ego's co-founder and CTO, former Meta researcher, said he would just close his eyes and rush forward:

I strongly recommend you to give it a try.

Many scholars also said that it is useful:

Search relevance is up to 11 times higher

As mentioned at the beginning, in order to test the performance of Lumina, the research team open-sourced a benchmark.

Using GPT-4 as the "judge", the relevance and accuracy of Lumina Base (basic search mode), Lumina Recursive (recursive search mode), Semantic Scholar and Google Scholar search results were compared.

Direct testing evaluates each search engine'sFirst 10 search results, the comparison is as follows:

In basic search mode (Lumina Base), Lumina search results are 4.8 times more relevant than Google Scholar and 8 times more relevant than Semantic Scholar, based on a generated dataset of approximately 2,470 queries.

In recursive search mode, Lumina further improves the relevance of search results, which is 6.8 times higher than Google Scholar and 11.3 times higher than Semantic Scholar, also based on 2,470 query data sets.

In addition, Lumina can consistently provide 2-3Highly correlatedAs a result, by comparison, only 50% of Google Scholar searches provided 1 highly relevant result, while only 30% of Semantic Scholar searches met this standard.

In other words, Lumina can find research results that would otherwise be hard to find.

The team also says that using the Lumina API, which will be released soon, and a simple recursive script, they can get relevance 11 times higher than Google Scholar.

In addition to basic search, Lumina also providesAI OverviewFunction.

For example, if you search for the keyword "machine learning", it will cite the retrieved papers summarizing machine learning:

Clicking on an underlined keyword in the AI ​​overview will trigger a further search for that keyword:

Below the AI ​​overview is a summary of the charts of search paper research results:

Created by a three-person team

Lumina is backed by a development team of only three people and has received investment from YC.

Co-founder and CEO Mehul Chadda graduated with a bachelor’s degree in materials engineering from the University of Wisconsin-Madison. He was a former manufacturing engineer at CAMECA, a high-precision scientific instrument company, and was responsible for the research and development of atomic probes.

Co-founder and CTO Akhilesh Sharma graduated with a bachelor’s degree in mechanical engineering from the University of Illinois at Urbana-Champaign. He was previously the co-founder and technical director of the social media platform Postpress, and also worked as a cloud technology consultant for the American technology consulting firm Neudesic.

Co-founder Ishaan Kapoor graduated from UCLA with a degree in statistics and is interested in deep learning, linear algebra, and large language models.

In addition to the new academic search engine, the team also developed an AI paper interpretation tool:Lumina-chat。

Click "Legacy System" in the upper right corner of the Lumina homepage to jump to use it.

After opening, it looks like this. You can upload PDF to interact with AI to interpret the paper:

Interested children's shoes can try it~