Gemini's long context and multimodal skills

0:00 Long context and multimodal

What if you could ask one question across forty documents and two hours of meeting recordings, and get an answer that points to the exact page and minute? That is what long context and multimodal models make possible. In this lecture you will learn how to use Gemini this way, and how to keep the answers verifiable.

0:25 Why this matters

Why does this matter? Because the information you need is usually buried in long material. A two hour webinar, a three hundred page report, a folder of forty documents. Reading all of it takes hours, and skimming misses things. Long context and multimodal models let you ask precise questions across everything at once. Think of it like a research assistant who has read the whole library, and who can point to the exact shelf and page for every answer, so you can check.

1:01 What makes it possible

Gemini models are built to be natively multimodal, accepting text, images, audio, video and PDFs, and many support very large context windows, on the order of a million tokens for several models, depending on model, product and plan. In September twenty twenty six, Gemini three point five Flash, launched at Google I O in May, is the default in the Gemini app, with Pro tier and Deep Think reasoning options on higher plans. Check the picker for current names and limits.

1:36 The long-context pattern

Here is the pattern. Label every source, Doc A is the annual report, Doc B the strategy deck, Video C the town hall. Give a targeted task, such as identify every commitment leadership made about expanding into Saudi Arabia. Ask for where each came from, the page or the timestamp, and whether other sources agree or conflict. Ask for not found in materials when something is missing. And request a table plus three open questions.

2:09 Why targeted beats summarise everything

Why targeted questions? Because a long context is not perfect recall. Summarise everything flattens nuance and can skip the one clause that matters. Specific questions surface specific facts, with locations you can check. Ask five precise questions instead of one vague one, and you will get five answers you can actually verify.

2:32 Audio and video

For audio and video, ask for decisions and actions with owners and timestamps from meetings. Ask for webinar summaries with timestamps and three quotable lines in exact words. With consent, ask which objections customers raised on calls, quoted with timestamps. And extract chapters and blog outlines from your own videos. Always spot check timestamps and quotes before using them.

2:58 Images and scans

Gemini also reads screenshots, charts, handwriting and scanned documents. When numbers matter, use the describe first pattern. List what you see, then answer, then flag anything uncertain. It catches misreadings before they become conclusions. It takes a few extra seconds and saves you from building an argument on a misread number.

3:20 Gemini Notebook

For answers limited to a trusted set of sources, use Gemini Notebook, which was called NotebookLM until July twenty twenty six. Same product, same notebooks. It answers only from the sources you add, cites the passage behind each answer, and can generate study guides, briefing documents, mind maps and Audio Overviews, podcast style discussions of your sources. We go deeper in the next lesson.

3:48 Simple example

A simple example. Upload a forty five minute webinar recording you are allowed to use, and ask, list the five main recommendations with timestamps, and quote the speaker's exact words for each. You get five points with times. Click to twenty three minutes fourteen seconds and listen. The quote matches. Now you can share the five recommendations with your team in two minutes, and anyone can jump to the exact moment to hear it for themselves.

4:21 Worked example: Abu Dhabi onboarding

An operations manager in Abu Dhabi must brief a new regional director using forty documents and two recorded leadership meetings. She labels every source, asks Gemini for commitments, risks and open decisions with pages and timestamps, and builds a Gemini Notebook with the same sources so the director can ask grounded follow up questions. Before sending, she checks ten random citations and fixes one timestamp that pointed to the wrong moment. The director's first week looked very different from a normal onboarding. Instead of reading forty documents, she asked the Notebook focused questions and followed the citations to the passages that mattered. She told the operations manager that the briefing table plus the Notebook saved her days, and more importantly, every answer could be traced to a source she could check.

5:18 Pitfalls

Four pitfalls. Summarise everything prompts. Using exact quotes without checking the recording. Uploading recordings without participants' consent. And assuming long context means nothing is missed. Success looks like answers with checkable locations, spot checks that pass, and hours of manual review saved.

5:36 Try this now

Try this now. Gather three documents or one long recording you are allowed to use. Label each source clearly in your prompt. Ask three targeted questions, not summarise everything, and require the page or timestamp for every point, plus not found in materials when something is missing. Then check three of the answers yourself by opening the page or jumping to the time. Note how accurate the locations were. That tells you how much you can rely on this workflow for your own material.

6:13 Key idea

Here is the key idea to take away. A long context gives the model coverage, the ability to look at everything, but not perfect recall of every detail. Two habits close that gap. Ask for locations, pages and timestamps, so every answer is checkable. And ask specific questions instead of requesting one giant summary. Coverage plus locations plus precise questions is what turns a long context model into a reliable research partner.

6:44 Watch me do it, part 1

Let me run the Abu Dhabi briefing. I upload three sources and label them in the prompt. Doc A, the annual report. Doc B, the strategy deck. Video C, the chief executive's town hall recording. Then the targeted question. Identify every commitment leadership made about expanding into Saudi Arabia, what was said, where, with page numbers for documents and timestamps for the video, and whether the sources agree or conflict, and write not found in materials if something is absent. The table comes back with seven commitments, each with a location, and one conflict. The deck says two stores in year one, the town hall says three.

7:31 Watch me do it, part 2

Now I verify. I click the first timestamp and jump to twenty three minutes fourteen seconds. The chief executive says three stores, matching the table. I open page forty one of the report, which matches too. The third location points to the wrong minute, the quote is actually about ninety seconds later, so I correct it. Then I create a Gemini Notebook with the same three sources and share it with the new director. She asks, what did leadership say about hiring in Riyadh? The answer cites the exact passage, and she can click straight to it.

8:13 Recap and next step

Recap. Gemini can reason across long documents, audio and video in one go. Label sources, ask targeted questions, demand pages and timestamps, and spot check. Use Gemini Notebook when answers must stay within your sources. Your next step: apply the long context pattern to a recording or three documents you are allowed to use, and verify three cited points.