Wednesday, September 17, 2025

SearchResearch Challenge (9/17/25): New agentic ways to get regular updates on a topic

Having a Research Assistant working for you is great... 

Image by Gemini showing [a scholar writing in front of a calendar], capturing the
need to do periodic updates on research topics. I love that it's SO old-fashioned.


... especially if you can get them to work on a regular schedule and produce useful, insightful summaries of things you should know.  

The advantage of a human Research Assistant (RA) working on your behalf is that they deeply understand what you're interested in and how that shifts and changes over time.  

The disadvantage of a human Research Assistant is that they're really expensive.  They're wonderful, but very few people can afford their own RA.  

We've discussed before about having regular Google Alerts that will constantly monitor the internet for items your cached searches will turn up.  

We've also talked about Google Scholar Alerts (same idea, except just within Scholar) and YouTube AlertsSemanticScholar has the same concept--a "standing query" that can run on a regular schedule, updating you with the latest results.  There are several other services that do similar things, discussed in this SRS post from 2023.

It should then come as no surprise to find that many of the current AIs have similar ideas.  ChatGPT and Gemini both have ways to set up a periodic alert-scan that will send you an update on whatever your topic-of-interest is. (Perplexity, Llama, and Claude don't currently support this.)  Basically, you set up an agent to periodically run a compare-and-contrast search for you on your topic.  

They're called different things by different companies, but whether you call them GPTs (OpenAI), GEMS (Gemini), subagents (Anthropic), characters (Meta)... or whatever... they all need a bit of setup.  

But setting them up is a bit like anticipating the chat you'll need to have.  Here are some tips about setting up these regular summaries.  

ChatGPT (Sep 16, 2025)

Once you've got your weekly task setup, you can click on the 3 dots to edit it.  You'll want to give it a reasonable name and clear instructions.  



Note that I've changed the default date/time of summary.  My prompt says what to search, the time scope (last 7 days), the topic (human sensemaking), with instructions about what kinds of sources I'm interested in (peer-reviewed papers...),  and how to create the summary (concise weekly summary with title, venue/source...).  

Here's a similar way to do this on Gemini: 

Gemini (Sep 15, 2025)  



(The instructions are clipped in this screencap.)  

In all cases, you want your prompt to follow all of the guidelines you know from setting up "regular prompts" (e.g., be specific, be directive, say what you want and what you don't want).  

Sample output looks like this (sent as email):  

ChatGPT output  (sent from OpenAI <noreply@tm.openai.com> on the sensemaking topic)



The output from Gemini is fairly similar.  I don't yet have enough experience with them to say which I prefer; I'll let you know if I find a big difference. 

Important:  How to manage your automated tasks.  

Gemini: Open a new Gemini window and go to the settings icon.  (Should be on the bottom left--looks like a gear icon.)  Click on "Scheduled Actions" to see all of your automated tasks.  

ChatGPT:  Very similar-go to Settings (bottom left). Click on the person icon (or your initials), then Settings, then "Schedules."  

From these pages you can edit the task or delete them.   



SearchResearch Lessons 

Upshot: Here's another method for keeping up-to-date on a topic area.  Now that we've got so many ways to do this, be careful to monitor your total feed--it's very possible to setup so many feeds (so many newletters!) that you can be overwhelmed.  

1.  Keep track.  I keep track of how much time I'm spending reading on a topic so I can be aware of how invested I am in a particular domain.  Every so often (roughly quarterly) I take an hour and assess how much time I'm spending on a given topic.  If it's too much, that's when you decide to delete something from your personal feed.  

2. Tune your instructions. When setting up your weekly prompt, be sure to write it out as clearly as you can.  Use your knowledge from doing regular prompting to inform your summary agent prompt.  


Keep searching!   (Just don't overdo it...)  



I changed the prompt to include "modern researcher writing on a dual monitor computer"



20 comments:

  1. Hello & Thank you, Dr. Russell.

    This time was not a Challenge nor an Answer. This is a SearchReSearch tools update. And it's very helpful.

    I was looking some weeks ago information about those gems and just now understanding them

    ReplyDelete
  2. Yes, the Gems aren't very clearly explained. They're basically Gemini + a special system prompt that you can add + any documents you might want to load up (making it more RAG-like).

    ReplyDelete
    Replies
    1. Thank you, Dr. Russell.
      What RAG-like means?

      Also wondering AI results with this new similar alert are better than regular alerts? This new tool searches all the available sources instead of a single one like in YouTube or Google Scholar, right?

      In the prompt you could add to search in other languages? In your search maybe it's not helpful as you already mentioned the trusted sources but in cases in which we don't have something as specific?

      Thank you, Dr. Russell!

      Delete
    2. " Retrieval Augmented Generation (RAG)"
      https://arxiv.org/html/2409.15566v1
      https://www.mygreatlearning.com/blog/retrieval-augmented-generation/
      "RAG stands for Retrieval-Augmented Generation. It is an AI framework that combines two main processes to produce more accurate and contextually relevant responses from large language models (LLMs). The process involves retrieving relevant information from an external knowledge base and then augmenting the LLM's internal knowledge with this retrieved data before the final generation of a coherent answer."
      Dan's explanation will be better...

      Delete
    3. Thank you, Remmij :)
      Good morning everyone

      Delete
    4. Remmij & everyone I searched after reading Remmij and found this video with [What is RAG LLMs for dummies on YouTube] I'm sure just RAG LLMs was enough

      From IBM. A very easy to understand video
      https://youtu.be/T-D1OfcDW1M?si=6vAEcMqOnxRIrtch

      Delete
    5. thanks Ramon, I found the IBM video helpful
      these two popped up in the right column - added detail... nice search work.
      https://youtu.be/VSFuqMh4hus?si=s66eL1a6um2QSyAd
      https://youtu.be/_HQ2H_0Ayy0?si=KleISG1_py2vXrkC

      Delete
    6. Good sleuthing! You folks are truly SRS stars.

      Delete
    7. Thanks Dr Russell & Remmij.

      I just read that now people can share Gems. Yours, Dr. Russell, is an example of those gems that can be shared or are different? In the blog, Google example says: Meal planner but that is all. We can't see what that does

      https://blog.google/products/gemini/sharing-gems/

      Delete
    8. Ramon, found this off your link... found it helpful for the things I attempt - thanks (the specificity can be good, but I miss the random 'noise fill-in' the would occur, often taking images in new, unexpected directions)
      https://blog.google/products/gemini/image-generation-prompting-tips/
      https://gemini.google/overview/image-generation/
      https://www.nanobanan.ai/?gad_source=1&gad_campaignid=22996401779&gbraid=0AAAABBTPlPaZddcShhkGrnjvOBILRDuc1&gclid=Cj0KCQjw_rPGBhCbARIsABjq9cfQbRK-hitv2G5lepw-jIf7-FmTz3WRZftHaLVJeIpPRmfAnzpejggaAmCFEALw_wcB

      Delete
    9. Thanks Remmij. I didn't know the link of nano banana. Nor that we could create without using Gemini or with credits. I checked the $ and it's on € I wonder if it's available worldwide. I don't think I need to buy credits. I still haven't tried to create an image

      Have you tried buying credits?
      Did you accomplish the goal you wanted?

      Delete
    10. I have not... being a bit of a bumbler I think I will wait a while before wading into nano banana & currency exchanges :)
      not a goal, rather an apparition...
      https://i.imgur.com/SWZ9a4E.jpeg
      https://gemini.google/overview/image-generation/
      fwiw - AI introduces a new dimension of portrait photography:
      https://i.imgur.com/hOFG4ib.jpeg

      Delete
  3. seems a tad sexist on Gemini's part that all the RAs are female...
    I'm sure it's just an 'unanticipated consequence'/coincidence...
    (I'm being mildly sarcastic)
    https://i.imgur.com/YkxTeJ8.jpeg

    ReplyDelete
  4. ...just for grins - AI_RA...
    https://i.imgur.com/ggt2gWu.jpeg

    ReplyDelete
  5. Just to celebrate the 6th Anniversary of the Joy of Search !

    Congratulations Dr. Russell 😊🙏🤗

    ReplyDelete
  6. The "Disembodied Hand" of AI (in any of Its assorted forms)
    my latest search attempt in the Time of AI... no stars about...
    but "SCAT" abounds... Search Caching Analysis Technology
    https://i.imgur.com/DPpEHoR.jpeg
    revised:
    https://i.imgur.com/gEWbVMq.jpeg

    ReplyDelete
  7. am starting to think this is a "real" thing, but still need to verify - will check Amazon for book & author... AI might be yanking my chain? (but that would mean it has a sense of humor - infantile & twisted as it might be in this current form, but not malicious... yet)
    https://i.imgur.com/6Shz5nN.jpeg

    "Russ M. Daniel author
    While a widely known author named Russ M. Daniel could not be identified, a search revealed two prominent authors with similar names: Russell M. Stendal and Daniel M. Russell.
    Russell M. Stendal
    A missionary and author who was kidnapped by Marxist guerrillas in Colombia in 1983.
    Key book: Rescue the Captors, which describes his kidnapping and his work with the same rebels.
    Other works: He has written numerous other books, many transcribed from his radio sermons, and edited the Jubilee Bible.
    Ministry: He runs the Colombia Para Cristo organization, which uses radio broadcasts to share messages of peace in Latin American conflict zones.
    Daniel M. Russell
    A former Google senior research scientist and a specialist in search quality and human-computer interaction.
    Key book: The Joy of Search: A Google Insider's Guide to Going Beyond the Basics, which provides guidance on advanced search techniques.
    Online education: He is also known for developing the "Power Searching with Google" online course, which millions have taken.
    Career: He has worked at other major tech companies including Apple, Xerox, and IBM. "

    Based on his SearchReSearch blog and other activities, Daniel M. Russell covers topics related to effective online research, the psychology of search, the user experience of information, and the integration of artificial intelligence with human inquiry. He refers to himself as a "search anthropologist" to describe his focus on the human side of interacting with information.
    Key topics include:
    Search challenges and solutions: The blog regularly features "SearchResearch Challenges," where Russell poses an intriguing question and then explains the search tactics and advanced operators required to find the answer.
    Search skills and techniques: Russell teaches methods for enhancing search abilities, such as how to effectively use search operators and evaluate the credibility of sources. This includes promoting foundational skills like using Ctrl+F to find text on a page, which he notes many people don't know.
    Sensemaking and information foraging: Russell investigates the cognitive and behavioral aspects of how people find, understand, and use information in large, complex datasets.
    Human-computer interaction (HCI): He writes about how users interact with technology to find and process information. His work draws from his extensive background in HCI research at companies like Google, Apple, and Xerox.
    User experience (UX) and search quality: Having worked on "Search Quality and User Happiness" at Google, his writing often touches on the user's overall experience with search engines and how it can be improved.
    Teaching search literacy: Russell's blog and other materials are designed to enhance web literacy by teaching effective research strategies. This educational mission is also reflected in his "Power Searching with Google" online courses.
    The future of search and AI: He explores how humans and artificial intelligence can work together to find information more effectively, a topic he has called "amplifying human intelligence".
    Observational studies: He often shares insights from his field studies and interviews with search users, including his observations of how different generations and types of users conceptualize the search process.

    ReplyDelete
  8. searched SCAT:
    "Based on search results, there is no recognized acronym for "SCAT" that stands for "search caching analysis technology". The results indicate that SCAT is an acronym for several different concepts, but not the one requested.
    However, search caching analysis is a legitimate computer science and web technology concept.
    Concepts of search caching and analysis
    Search engine caching: Search engines cache web pages to provide faster query responses for users. An analysis of this caching might involve evaluating how long results are cached versus the freshness of the data.
    Website and browser caching: Websites and browsers use caching to store local copies of data like images and scripts. Analyzing this caching improves website performance and user experience.
    System and network caching: Caching can be implemented at various levels, such as in-memory, virtual memory, and on storage controllers. The analysis of these caches is used to optimize system performance and reduce latency.
    Large Language Model (LLM) caching: With the rise of LLMs, new caching architectures have been developed. For example, "semantic caching" uses embedding vectors for approximate searches, and analysis of this technology helps optimize performance.
    Common meanings of SCAT
    The search results provide multiple definitions for the SCAT acronym in different fields.
    Sneak Circuit Analysis Tool (SCAT): A software program used by engineers to automatically identify potential flaws in electrical circuit designs.
    Systematic Cause Analysis Technique (SCAT): A method used to investigate and identify the root causes of incidents or failures in various industries.
    Shoreline Cleanup Assessment Technique (SCAT): A process used in oil spill response to document the extent of shoreline oiling to guide cleanup operations.
    Signaling Collection and Analysis Tool (SCAT): A software application for parsing diagnostic messages from cellular basebands."


    ReplyDelete
    Replies
    1. lost in the shuffle? maybe AI has a sense of humor?
      https://i.imgur.com/kx3yUJA.jpeg

      Delete
  9. pt.2
    "Search caching analysis involves studying user query logs to understand query locality (how often queries are repeated and shared) to design effective caching strategies that improve search engine performance. Key aspects of analysis include cache hit rate, which measures how often cached results serve requests, and the Zipf distribution of queries, revealing that a few queries are very popular while most are unique. Analysis guides the implementation of different cache types, such as static (for popular queries), dynamic (for bursty queries), and prefetching (for future user needs), to reduce response times, lower hardware costs, and optimize resource allocation.
    Key Aspects of Search Caching Analysis
    Query Locality and Zipf Distribution:
    Analysis of search engine logs shows that query frequencies often follow a Zipf distribution, meaning a small number of queries are highly popular and can be effectively cached, while most queries are unique and less suited for server/proxy caching.
    Cache Hierarchy:
    Analyzing query patterns helps determine the optimal location for caches (e.g., user-side, proxy, or server-side) to maximize hit rates and leverage different types of locality.
    Caching Policies:
    The analysis informs the development of caching policies, like the probability-driven cache (PDC), which uses a probabilistic model of user behavior to prioritize what to store.
    Prefetching Strategies:
    By understanding user behavior, especially the common words used in queries, analysts can predict future queries and prefetch relevant content to increase cache hit ratios and serve users faster.
    Metrics for Analysis
    Cache Hit Rate:
    The primary metric, indicating the percentage of requests served from the cache rather than requiring a full recomputation.
    Response Time:
    The speed at which users receive search results, which caching significantly improves by serving results from faster, local storage.
    Hardware Resource Savings:
    Analysis demonstrates that effective caching reduces the computational load and hardware requirements for search engines.
    Types of Caches Analyzed
    Static-Dynamic Cache:
    Analyzes query logs to create a static part of the cache for popular, long-term queries and a dynamic part to handle short-term, bursty query behavior.
    Similarity Caching:
    Explores caching for content similar to, but not identical to, the requested query, evaluating user-oriented metrics beyond traditional hit ratios.
    Machine Learning Approaches:
    Modern approaches use machine learning on query logs to identify features that improve the accuracy of caching decisions and boost cache hit rates"

    ReplyDelete