Gemini Notebook (NotebookLM) Size Limits: Document, Audio, and Source Constraints

Google's Gemini Notebook, formerly known as NotebookLM, restricts individual sources to 200MB or 500,000 words, with each notebook holding up to 50 sources on the standard tier, according to Google's official documentation. While Google renamed the service to Gemini Notebook in July 2026, the core ingestion architecture and file boundaries remain identical. When researchers or automated pipelines attempt to feed dense books, multi-hour audio recordings, or comprehensive project archives into a single notebook, import fails silently or returns unhandled upload errors.

Every upload encounters a dual-ceiling architecture. Google evaluates files along two separate dimensions: raw storage volume in megabytes and linguistic density in words. Reaching either boundary triggers an immediate import rejection. A plain text file containing 450,000 words may occupy less than 3MB on disk, yet it sits right against the word cap. In contrast, an uncompressed archival scan with high-resolution graphics may reach 195MB while containing only 8,000 words of text. Both files are valid inputs, but exceeding either threshold aborts the ingestion process.

The specifications table below summarizes the exact limits across all supported source formats verified against Google Support documentation in October 2026:

• Google Gemini Character Limits: Prompt Constraints and Large File IndexingThe Gemini character limit restricts direct prompt pasting in the web interface, creating friction for users attempting...
• Google NotebookLM Character Limit: Source Constraints, Word Caps, and WorkaroundsGoogle NotebookLM restricts individual sources to 500,000 words, which translates to roughly 2.5 to 3 million...
• Google NotebookLM Message Limits: Chat Query Caps and Audio QuotasGoogle NotebookLM message limits encompass prompt input length constraints (roughly 2,000 to 4,000 characters per...
• Perplexity Token Limits: Document Size, Query Caps, and Large Corpus WorkaroundsThe Perplexity token limit defines the maximum token capacity Perplexity can parse from attached documents (typically...
• ChatGPT Canvas Limits: File Sizes, Document Length, and WorkaroundsChatGPT Canvas limits documents to practical thresholds of roughly 4,000 lines of code or 25,000 words, running in an...
• Gemini CLI Token Limits: Context Windows, Piping Caps, and File ArchitectureA Gemini CLI token limit is the operational ceiling on input prompt tokens, terminal pipe payloads, and context window...

More on this subject: AI Agents: General Guides (190 guides)

Audio file handling introduces a distinct mechanical dynamic that trips up many research teams. Competitors frequently overlook how raw audio file sizes and transcription word counts interact, leaving users confused when audio uploads fail unexpectedly. Google allows direct uploads of audio files up to 200MB. Supported audio extensions include MP3, WAV, M4A, AAC, OGG, FLAC, OPUS, WMA, AMR, and 3GP. Once uploaded, Gemini Notebook does not store or process the raw audio during user conversations. Instead, an automated Gemini speech-to-text pipeline transcribes the audio into a written transcript at the moment of ingestion. That generated transcript becomes the text source saved into the notebook. This two-stage pipeline creates an asymmetry between file size and duration. The primary limitation on audio uploads is almost always the 200MB file size ceiling rather than the 500,000-word transcript ceiling.

Uncompressed high-fidelity audio formats consume substantial storage per minute of recording and reach the file size ceiling quickly, yielding only a fraction of the allowable transcript length. In contrast, voice-optimized compressed audio formats require less storage per hour of speech, allowing multi-hour meeting recordings or interviews to fit comfortably within the upload threshold. Because conversational speech proceeds at an ordinary cadence, the resulting transcript easily stays within the textual capacity of the notebook. The active bottleneck is therefore file compression and format selection, not the textual capacity of the notebook.

Gemini Notebook organizes research material within individual notebooks, each containing an isolated set of grounded sources. On the standard tier, Google caps each notebook at 50 sources. This boundary represents the most common operational ceiling encountered by researchers, analysts, and project teams.

Google provides paid plan upgrades that expand notebook quotas and total source capacity. The breakdown below details the published source caps across consumer and workspace tiers verified against Google's plan availability guide in October 2026:

Upgrading to an advanced plan increases the number of sources permitted per notebook, moving the limit from 50 up to 600. However, an upgrade does not increase the per-source file size or word count ceiling. A subscriber on the highest Ultra tier remains bound by the exact same 200MB file size limit and 500,000-word cap per source as a user on the standard tier.

Collaboration rules enforce another important constraint: sharing a notebook does not change the source limit for any collaborator. If a project lead on a Pro account creates a notebook containing hundreds of sources and shares edit access with colleagues on standard accounts, those collaborators can interact with and query all sources within that shared notebook. However, the collaborators cannot add sources beyond the notebook owner's plan quota, nor does participating in a large shared notebook expand the 50-source limit on their personal notebooks.

Query Research Archives Beyond Document and Source Limits

Keep entire project libraries, multi-hour transcripts, and dense PDF sets in a shared Fast.io workspace. Enable Intelligence Mode to index files for semantic search, and connect AI assistants through the Fast.io MCP server. Monthly plans start with a 14-day free trial (credit card required).

Start 14-Day Trial

When an import fails in Gemini Notebook, the application displays a generic error message rather than specifying the exact constraint violated. Understanding the root causes of ingestion failures allows teams to resolve issues quickly.

Google documentation identifies three primary causes of direct import failure: the file exceeds 200MB, the file exceeds 500,000 words, or the file contains copy-protection locks. Beyond those primary triggers, format-specific limitations account for the majority of rejected uploads.

The following checklist identifies common failure triggers and concrete remediation steps:

• Copy-Protected and Password-Encrypted PDFs: Many academic journals, legal filings, and financial statements apply Adobe Acrobat digital rights management or security permissions that disable text extraction. Gemini Notebook cannot bypass encrypted PDF containers. To resolve this, decrypt the document using an authorized utility like qpdf --decrypt input.pdf output.pdf before uploading.
• Image-Only Scanned Documents: If a PDF consists of raw scanned bitmap images without an embedded OCR text layer, Gemini Notebook may upload the file without generating usable citations or reject the file as empty. Process scanned documents through an OCR pre-processing tool before ingestion.
• Spreadsheets Exceeding Token Limits: Uploading a large CSV or Google Sheet with dozens of columns and thousands of rows causes silent truncation or immediate rejection once it hits the 100k tokens ceiling. Isolate the specific columns and summary tables required for your analysis rather than uploading raw production database dumps.
• YouTube Video Age and Transcript Availability: YouTube video imports rely entirely on pre-existing video captions. Videos published within the past seventy-two hours frequently lack indexed automated captions, resulting in import errors. Private videos, unlisted videos with disabled captions, and videos exceeding 500,000 words in their transcript will fail.
• Presentation Slide Counts: Google Slides decks containing more than 100 slides trigger a hard validation error. Split expansive slide decks into two smaller files before importing.
• Audio Without Spoken Content: Audio recordings containing music, ambient office noise, or long stretches of silence fail ingestion because the speech recognition model cannot extract semantic anchor text.

qpdf --empty --pages input_large_report.pdf 1-150 -- chapter1.pdf

ffmpeg -i interview_master.wav -vn -ac 1 -b:a 64k interview_compact.mp3

When research projects outgrow the 50-source boundary or require cross-corpus analysis over hundreds of documents, researchers face a difficult structural decision. Dividing a cohesive project into dozens of separate notebooks fragments synthesis, while manually summarizing documents to fit under word caps introduces human error.

Teams typically explore two common workarounds before seeking dedicated workspace infrastructure:

• Local Folder Archives: Maintaining directories of markdown and text files on a local drive gives researchers unlimited file storage, but lacks collaborative sharing and automated retrieval. Without an index, users must write custom search scripts or feed individual files into LLM contexts manually.
• Standard Cloud Storage: Keeping files in consumer cloud storage like Google Drive or Dropbox avoids local disk constraints. However, standard cloud drives treat documents as passive blobs. They do not index contents for semantic question-answering or provide grounding citations to connected assistants.

The alternative is an intelligent workspace like Fast.io, which pairs persistent file storage with automatic retrieval-augmented generation. Instead of attaching static files directly to an isolated notebook container, the entire research corpus lives in an organization-owned workspace where documents are indexed upon arrival.

The workflow for managing large research archives operates through three core capabilities:

• Unified Ingestion via Cloud Sync: Research teams can import documents directly or connect existing folders from Google Drive, Dropbox, Box, and OneDrive via Cloud Sync. Synchronization operates one-way or two-way, on a schedule or on demand, preserving existing file folder hierarchies.
• Intelligence Mode for Automatic Indexing: Turning on Intelligence Mode in workspace settings automatically indexes every document, spreadsheet, PDF, and transcript. Fast.io generates a hybrid search index combining full-text keywords, semantic embeddings, and metadata values, eliminating the need to deploy and manage a standalone vector database.
• Remote MCP Server for AI Assistants: External AI clients (including Claude Desktop, Claude Cowork, coding agents, and custom assistants) connect to the workspace using Fast.io's remote Model Context Protocol server. Using Streamable HTTP at https://mcp.fast.io/mcp/tools (or https://mcp.fast.io/mcp/code for developer agents), assistants query the indexed corpus directly. The assistant searches and retrieves relevant grounded excerpts with source citations on demand, leaving vendor upload limits unaffected while providing instant access to massive document libraries.

Collaborative research also benefits from per-file version history and a detailed activity log. When team members or autonomous agents update documents, extract structured fields, or add annotations, every change is tracked and previous iterations remain recoverable.