Google rebranded NotebookLM. It iis now called: Gemini Notebook.
Gemini Notebook is a closed RAG system. You upload sources, and Gemini Notebook generates responses grounded narrowly in the specific data in your sources. This grounded approach reduces the risk of hallucinations.
Gemini Notebook can also be considered a document summarization tool, as the response involves synthesizing data and answering questions based on the content of the sources.
Sources can include complex documents, and prompts can have large context windows of up to one million tokens (500,000 words per source).
Inputs
Multi-modal source input options that include Google Drive (e.g. Docs and Sheets), files (e.g. PDF, DOCX, and CSV), audio/video (e.g. MP3, MP4, and WAV), and images (e.g. GIF, JPG, and PNG), as well as searching the web and Deep Research.
Outputs
Studio outputs include Audio and Video Overviews synthetic podcasts, Mind Maps, Reports, Flashcards, Quizzes, Infographics, Slide Decks, and Data Tables.
The new Audio Overview feature from your sources transforms data into engaging “Deep Dive” discussions with one click.
Gemini Notebook provides clear citations for its work, showing the exact quotes from your sources.
A closed RAG system is an AI setup that combines an LLM with a strictly private, offline knowledge base.
It allows the AI to answer questions using only your specific documents, preventing it from hallucinating or accessing the public internet. This provides secure, fact-based answers grounded purely in your proprietary data.
Vector Database: These vectors are safely stored in a closed-off database.
Retrieval Phase: When a user asks a question, the system converts the query into a vector and searches the database to find the most contextually relevant document chunks.
Generation Phase: The retrieved chunks are sent to the LLM alongside the user's prompt. The LLM is instructed to only use this information to formulate its response.
Why Use a Closed System?
Organizations typically use closed RAG systems instead of public models or the open web for a few reasons:
Strict Data Privacy: Sensitive corporate, legal, or medical data never leaves the internal, secured environment.
No hallucinations: Because the model is heavily restricted to the provided documents, it cannot invent "facts" or pull in unverified outside information.
Source Attribution: The system can cite the exact documents and page numbers used to generate the answer, building user trust.
Real-World Examples of How Gemini Notebook is Used:
Academic Study: A student studying for a university exam can upload 20 class PDFs and textbook chapters to Gemini Notebook and ask the AI to "create a study guide with definitions and a practice quiz based only on these documents".
Research and Writing: An author or researcher can compile a notebook of old journal articles, interview transcripts, and prior notes. They can then instruct it to "summarize the interactions between historical figures X and Y based on these sources.
Medical Practice: A healthcare provider or coach can upload clinical studies and generate patient-friendly briefs that explain study findings and key FAQs in simple terms".
Travel Planning: A traveler can upload local blogs, restaurant guides, and hotel brochures to ask the AI to "draft a 3-day itinerary staying strictly within our travel budget.
Auto-Generated Podcasts: You can use its Audio Overview feature to transform your documents into a two-host, podcast-style discussion summarizing your source material.
While you can name the Notebook itself to help the AI maintain focus, you can explicitly set global guardrails and instructions for the chat by opening your notebook and clicking the Configure Chat settings icon (which looks like sliders/controls) in the top-right corner of the chat panel.
Setting Up Global Chat Guardrails
Click Configure Chat settings icon in the chat bar.
Under the conversation style options, select Custom
Paste your encompassing title, focus, or strict boundaries into the text box (up to 500 characters).
Examples of Guardrail Prompts:
Instead of just a simple title, use a clear framework within that setting to ensure the AI stays on track:
Example 1 (Medical & Legal Compliance):
"Role: Strict Fact-Checker. You are restricted to analyzing the clinical trials in this notebook. Do not extrapolate, give medical advice, or assume outcomes. If a drug interaction isn't explicitly noted in these exact files, reply 'Not found in source data'."
Example 2 (Product Strategy & Focus):
"Project Focus: Q3 Onboarding Friction. Every response must evaluate data strictly through the lens of user friction during account creation. Ignore marketing or sales data in the sources unless it directly impacts the signup drop-off."
Example 3 (Strict Formatting & Style):
"Rule: Concise Executive Summaries. Never exceed 200 words. Always lead with a 1-sentence bottom-line takeaway, followed by 3 bulleted supporting facts directly cited from the uploaded sources."
Project Naming
AI research workflows indicate that naming your notebook tightly around one central question acts as an initial guardrail.
Instead of naming a project "Marketing Documents," naming it "Should we launch Product X in the European Market?" forces the underlying Gemini model to interpret all mixed sources strictly through that single analytical framework.
Gemini Notebook can customize chats using multiple conditions simultaneously. The system does not use a coding language or rigid syntax.
Because the "Custom Instructions" box allows for up to 10,000 characters, you can combine multiple logical constraints (such as persona, format, rules, and negative constraints) into a single, cohesive system prompt.
To successfully enforce multiple conditions without confusing the AI, structure your instructions using Markdown headers (###), bullet points (*), and bold text (**). This visual separation serves as clear semantic boundaries for the Gemini model.