Research That Remembers

Level 1 · Research That Remembers

1. Concept: Find it, with sources

   Ask a web-search AI about your customers, competitors, suppliers, grants and rules, and keep only the claims that come with a link.

   1. Term: Citations

      Footnotes that show which page, and which passage, a claim came from.

      See also Retrieval

      1. In plain words: A receipt for a fact

         A citation is the receipt for a fact. It points to the page the AI read, so you can check the claim yourself instead of taking its word for it.

      2. How it works: Search first, then write

         AI tools with web search look pages up first, then write an answer with links to what they read. Without search, the AI answers from memory, and a source it names may not even exist.

      3. Try it: Open three links

         Ask a web-search AI who your main competitors are, with a link for each one. Open three of the links and find the sentence on each page that backs up the claim.

      4. Watch out: Footnotes can be wrong too

         Footnotes cut down on made-up answers, but they don't end them. A 2024 Stanford study found that leading AI legal research tools, which look things up before they answer, still got 17 to 33% of answers wrong or wrongly cited.

         See also Staleness

   2. Term: Deep research

      A slower mode where the AI reads many pages for several minutes and writes a report with links.

      1. In plain words: A report, not a reply

         Instead of one quick search, the AI makes a plan, reads lots of pages and writes a report. It takes minutes instead of seconds, so start it and go refill your coffee.

      2. What it costs: Free to try, more if you pay

         Gemini's free plan includes Deep Research, and ChatGPT's free plan has a limited amount. Claude keeps its Research mode for paid plans, about $20 a month. Checked Oct 23, 2026; these limits change often.

      3. Watch out: Long is not the same as right

         A long report sounds sure of itself. Treat it like a first draft from a new hire: quick and useful, with every claim still needing its link checked.

         See also Citations

   3. Term: Research prompts

      Name the place, say what you need for each item, and tell the AI to admit when it can't find something.

      1. In plain words: Specific beats clever

         A good research prompt names your town, your kind of business and the facts you want for each item. Then it adds one rule: only include what you can link to.

      2. Try it: Borrow this one

         Try: "List lawn-and-snow companies serving Cedar Falls and Waterloo, Iowa. For each: services, any published prices and a link. Say when you couldn't find a price." Then swap in your own business and town.

      3. Watch out: Give it room to say no

         An AI that isn't told it can come up empty will sometimes fill the gap with a guess. Asking it to say when it couldn't find something is the cheapest fix in this whole class.

         See also A missed card means a guess

2. Concept: Save it, with dates

   Keep your best sources in one folder you control, with the date first in every file name.

   1. Term: Dated files

      Each good page becomes a PDF in one folder, named with the date first.

      See also Staleness

      1. In plain words: A binder with dates on every page

         Your saved sources are the binder. Every one gets two dates: when it was published or updated, and when you saved it.

      2. How it works: Ctrl+P, then one line in the log

         Press Ctrl+P, choose Save as PDF, and name it like 2026-10-23 USDA REAP rule.pdf so the folder sorts itself by date. Then add a line to your source log: title, link, published date, date saved and why it matters.

      3. Watch out: A saved page is a snapshot

         A saved page never updates itself. Treat a web page you add to Gemini Notebook the same way, as a copy from that day, and add it again to get the new version. Google Drive files are the exception, because Gemini Notebook keeps those in sync.

         See also Refresh the binder

   2. Term: Cloud or laptop

      The same binder can live in a cloud notebook or privately on your own laptop, and in class you build both.

      1. In plain words: Two homes for one binder

         In the cloud: Gemini Notebook (formerly NotebookLM), or a Claude or ChatGPT Project. On your laptop: AnythingLLM Desktop, a free app that keeps your files and its search index on your own machine.

      2. What it costs: Free to start either way

         Gemini Notebook's free plan holds 100 notebooks of 50 sources each. Claude's free plan allows 5 projects, ChatGPT's free projects hold up to 5 files, and AnythingLLM is free and open source. Checked Oct 23, 2026.

      3. How it works: Same files, same questions

         In class you load one folder into both and ask both the same three questions. Then you compare the footnotes, and how long each one takes to answer.

      4. Watch out: Private, with two catches

         On the laptop your files stay put, but a web search still goes out to the internet. And if you hook the app up to a cloud AI, the cards it reads go to that company.

         See also Privacy

   3. Term: Privacy

      Cloud tools keep what you upload, so public research goes in and private records stay out.

      See also Cloud or laptop

      1. In plain words: Who else can read this?

         Anything you upload to a cloud AI sits on that company's computers. Research about your market is usually fine; records about your customers and your money are not.

      2. How it works: The training switches

         Consumer plans may use your chats to train their AI unless you switch that off: Claude's privacy settings, ChatGPT's "Improve the model for everyone," and Keep Activity in the Gemini app. Gemini Notebook says it doesn't train on your data unless you send feedback. Business plans don't train on your data by default.

      3. Try it: Find the switch first

         Open your AI's settings and find the training switch before you upload anything. In class we flip them together in the first 15 minutes.

      4. Watch out: Keep these out of the cloud

         Keep these out of cloud tools: customer lists with personal details, employee records, Social Security numbers, bank or tax statements, passwords, and anything under an NDA. Private files belong in the laptop version with a local model, and passwords stay out of both.

3. Concept: Search by meaning

   The tool cuts your files into index cards, gives each card a meaning fingerprint, and hands back the cards closest to your question.

   1. Term: Chunks

      Paragraph-sized index cards cut from your documents.

      1. In plain words: Index cards

         A chunk is an index card cut from a document, about a paragraph long. The AI reads the few cards that match your question, not the whole binder.

      2. How it works: Cut with a little overlap

         AnythingLLM cuts every 1,000 characters or so, about a long paragraph, with a small overlap so nothing gets lost at the seams. Each card remembers which file it came from.

      3. Watch out: Card 37 can get missed

         Each answer reads only 4 to 6 cards, so a fact buried on card 37 can be missed, and a table can get split in two. For a short, crucial document, pin it in AnythingLLM or ask about it by name.

         See also Retrieval

   2. Term: Embeddings

      A meaning fingerprint: a list of numbers that captures what a card is about.

      See also Vector database

      1. In plain words: Meaning fingerprints

         Each card becomes a list of numbers, 384 of them with AnythingLLM's built-in model. Cards that mean similar things get similar numbers, even when the words differ, like free advice and no-cost advising.

      2. How it works: Same recipe for your question

         Your question gets a fingerprint made the same way. The computer then looks for the cards whose fingerprints sit closest to it.

      3. What it costs: Light work for an old laptop

         Free. The built-in model is about 25 MB and runs on the laptop's regular processor, no graphics card needed. A class binder of a few hundred cards should take a minute or less.

      4. Watch out: Not magic

         Small models miss sometimes. In a test binder we built for this class, "cheapest de-icer for parking lots" pulled a competitor price list ahead of the actual salt quote. Use the words your sources use, ask a second way, and look past the top card.

         See also Similar isn't true, or new

   3. Term: Vector database

      A filing cabinet sorted by meaning that stores your cards and their fingerprints and hands back the closest matches.

      See also Embeddings

      1. In plain words: A filing cabinet sorted by meaning

         Hand it a question and it hands back the cards that mean the most similar thing. Techies say "vector" because each fingerprint is a list of numbers.

      2. How it works: Closest first

         It compares your question's fingerprint with every stored one and returns the nearest few, each with a similarity score that says how close it is. In AnythingLLM the cabinet is called LanceDB, and it lives inside the app on your laptop.

      3. Try it: One folder, one question

         Drag a folder of saved sources into AnythingLLM and ask a question in your own words. Open the cited cards and see which ones came back first, and which ones you expected but didn't get.

      4. Watch out: Similar isn't true, or new

         It ranks by closeness of meaning, never by truth and never by date. An expired deadline comes back just as cheerfully as a fresh one, and its score isn't a grade for accuracy.

         See also Staleness

4. Concept: Ask later, with today's web

   Months from now, the AI pulls the best cards from your binder, checks today's web, and answers with both, dates showing.

   1. Term: Retrieval

      Looking up the best-matching cards first, then answering from them.

      See also Chunks, Vector database

      1. In plain words: An open-book test

         Retrieval turns a closed-book test into an open-book one: the AI looks in your binder before it answers. Techies call it RAG, short for retrieval-augmented generation.

      2. How it works: Question in, cards out

         Your question gets a fingerprint, the database returns the closest 4 to 6 cards, and the AI writes its answer from those cards, with footnotes. AnythingLLM's Query mode answers only from your files and tells you when it finds nothing relevant.

         See also Embeddings

      3. What it costs: It costs time, not money

         Finding the cards takes a blink. Writing the answer is the slow part: a small model running on these ThinkPads takes roughly 15 seconds to 2.5 minutes per answer, by our estimate, while cloud tools answer in seconds.

      4. Watch out: A missed card means a guess

         If the right card isn't found, the AI can't use it, and it may fill the gap with a confident guess. Tell it to say plainly when your files don't cover something.

   2. Term: Today's web

      A quick call to the live internet for anything that may have changed since you saved it.

      1. In plain words: The binder and the phone call

         Your saved research is the binder. A web search is the phone call to check what changed. A good answer uses both and shows the date of each source.

      2. How it works: Where the switches are

         Open your notebook in the Gemini app, where it can use your sources plus web search, or turn on web search in a Claude or ChatGPT Project. On your laptop, AnythingLLM's agent can search the web through DuckDuckGo with no key needed: start your question with @agent.

      3. Try it: The two-piles prompt

         Paste this: "Use my saved research and today's web, and cite each source with its date. My files win for facts about my own business; the newest official source wins for deadlines, rules and prices. If they disagree, show both dates and add the item to a Call to confirm list."

      4. Watch out: Searches leave the laptop

         Even in private mode, a web search goes out to the internet. Keep customer names and private details out of anything you search for.

         See also Privacy

   3. Term: Staleness

      Saved research expires, and search by meaning can't tell old from new.

      See also Dated files

      1. In plain words: Everything has an expiration date

         Deadlines, prices, program rules, laws and hours all expire. A note from March can be wrong by October, and nothing in the note will warn you.

      2. How it works: Ranked by meaning, not by date

         Search looks for the closest wording and ignores the date. In a test binder we built for this class, a March 2025 grant note came back first, ahead of an October 2026 update, just because its words matched the question better.

         See also Similar isn't true, or new

      3. Watch out: Same question, opposite advice

         That fictional 2025 note says to apply for USDA's REAP energy grant before you start a project. In real life, USDA rescinded its REAP funding notice on April 15, 2026, and a new rule effective October 16, 2026 says to build first and apply only after 12 months of energy data. REAP is for farmers and rural small businesses, so this one hits close to home in Iowa.

      4. Try it: Ask what changed

         Pick a saved answer about a deadline or a price and ask: "What in this has changed since it was saved?" Then open the new sources it finds and check their dates.

   4. Term: Which source wins

      Your files win for facts about your own business, and the newest dated official source wins for anything that changes.

      1. In plain words: Two simple rules

         Your files win for your own business: your prices, history, customers and decisions. The newest dated official source wins for deadlines, program rules, laws, permits and hours.

      2. How it works: Call to confirm

         When the two disagree, the AI should say so, show both dates, and put the item on a Call to confirm list. Then a person picks up the phone before anyone acts on it.

      3. Try it: Refresh the binder

         When you find a fresher source, save it with today's date and add it back, so the next question finds it. In Gemini Notebook, Fast Research and Deep Research can find new sources and import them for you.

         See also Dated files

      4. Watch out: A news story is not the rule

         Articles about a rule are handy, but the rule itself lives on the agency's own site, and for federal rules, in the Federal Register. For grants and laws, find the official page and check its date.