Combining AI Search and Search Queries: Practical Notes on Literature Search

Creating database search queries can be quite challenging to do on your own. That's when leveraging large language models like ChatGPT and Gemini can easily help you build these queries, so I highly recommend using them when you're struggling with literature searches.

Let me briefly summarize the overview of search queries and why they're necessary.

Do We Still Need Search Queries?

In this era of rising AI-powered search, are search queries still necessary?

After extensively using AI search tools, I believe search queries are still very much needed. There are several reasons, but the biggest one is the issue of accessing paywalled articles.

AI search tools in most cases can only read the full text of open access papers. This means there's likely to be bias in the papers that appear in search results, and we can't know whether we're truly accessing the paywalled papers we need.

Important papers are often not open access, so missing them due to these limitations can be critical.

Given the current trends in copyright lawsuits surrounding LLMs, I don't foresee a future where paywalled content can be easily entrusted to LLM searches, so I think this situation will remain largely unchanged.

Some tools like OpenEvidence have partnerships with journals to break through paywalls, but I think this is only possible because it's for limited use cases (= requiring partnerships with fewer journals).

Naturally, search queries can search for these articles as well, so they remain highly necessary.

Another issue is reproducibility.

AI search converts queries (questions from users) into multiple keyword searches, vectorizes text into numerical values to find similar items, and then goes through complex processes including AI agent selection to narrow down information based on literature content.

Honestly, users don't really understand this process. Therefore, even with repeated searches, it's unclear whether we've achieved comprehensive coverage.

Tools like Elicit try to make the information selection process as transparent as possible, but it's still quite difficult to actually adjust the internals well.

In contrast, keyword searches allow users to extract literature with reproducibility, enabling comprehensive literature review by trying multiple keywords while observing results. In fact, even after thinking you've thoroughly researched using AI search and citation analysis, keyword searches often reveal "I can't believe I missed so many matching results!"

For these reasons, I've recently become more conscious of properly combining keyword searches as well.

Using ChatGPT for Keyword Searches

LLMs like ChatGPT can of course help create search queries.

For example, when instructing it to create a search query for PubMed, a biomedical database, it constructs something like this:

Prompt: "Please create a search query that can comprehensively search PubMed for high-quality journals covering cell biology techniques and their review articles."

▼ChatGPT's answer

(
  Keyword1[tiab] OR Synonym[tiab] OR Related term[tiab]
)
AND review[pt]
AND ("2015/01/01"[dp] : "3000"[dp])
AND (
  "Nat Rev Methods Primers"[ta]
  OR "Nat Methods"[ta]
  OR "Cell Rep Methods"[ta]
  OR "Trends Cell Biol"[ta]
  OR "Trends Biotechnol"[ta]
  OR "Annu Rev Cell Dev Biol"[ta]
)

[tiab] searches titles and abstracts, while [ta] is a search tag for specifying journals. By confirming these details while building search queries, you'll better understand your own search content.

All you need to do is insert the topic you want to know about in the keywords. You can also select necessary journals and desired article types, so you can decide these through consultation as well.

Recently, I've been searching for reviews in high-impact journals and setting up alerts to get the latest information in one place. With the method introduced in the following article, you can even turn the latest information into podcasts to listen to or videos, which I highly recommend.

When you get too many hits from a search, you can consult with the LLM again on how to narrow it down, and as you progress while understanding the search queries, your search skills will improve.

Summary

Even in today's world where AI search tools have become convenient, traditional keyword searches remain important from the perspectives of accessing paywalled articles and reproducibility. However, since creating search queries is difficult, you can efficiently build high-quality search queries by leveraging LLMs like ChatGPT. By appropriately combining AI search and keyword searches, more comprehensive and high-quality literature surveys become possible. By understanding and improving search queries through dialogue with LLMs, your search skills will naturally improve as well.