Learn
How to use AI without uploading files every time
The upload ritual
Every AI conversation about your own work starts the same way. You open ChatGPT or Claude, drag in a PDF, wait for it to process, then ask your question. Tomorrow you need to reference the same document, and you do it all over again. The week after, you need it alongside two other files, so you upload three things. The model can only hold so much, so one of them gets truncated. You rephrase your question and try again.
This is the default workflow for millions of people, and it works just well enough to feel normal. But it does not scale. The more files you work with, the more time you spend managing uploads rather than thinking. The friction is low for a single document on a single day. Across weeks of work, across a library of hundreds or thousands of files, it becomes the thing that stops AI from being useful in the way it should be.
Understanding how LLMs use your files helps clarify why this happens. The model itself has no memory of what you uploaded yesterday. Each conversation is a blank slate. The upload is not building a relationship between the model and your content; it is a one-time injection that evaporates when the session ends.
Why uploads are the bottleneck
The problem is not that uploading a file is hard. It is that the upload model makes three assumptions that do not hold for real work.
The first assumption is that you are working with a small number of files. ChatGPT and Claude both impose limits on how many files you can attach and how large each one can be. A 50-page contract fits. A folder of project documentation does not. If your question requires context from multiple sources, you hit the ceiling fast. You start deciding which files are "important enough" to upload, which means you are doing the retrieval work that the AI should be handling.
The second assumption is that each conversation is self-contained. In practice, knowledge work is cumulative. A question you ask on Friday depends on context from Monday's meeting notes, last month's strategy document, and a Slack thread from three weeks ago. The upload model forces you to reassemble that context every time, as if your memory resets along with the model's. The page on how to stop re-explaining context to ChatGPT covers this pattern in detail.
The third assumption is that you know exactly which file contains the answer. Often you do not. You might know the answer is somewhere in your notes, or somewhere across your Google Drive and Notion workspace, but not which specific document holds it. The upload model has no search layer. It processes what you hand it and ignores everything else. That means the quality of the AI's response depends entirely on your ability to locate the right file before you even ask the question.
The cost of repeated uploads
Beyond inconvenience, the upload cycle has practical costs that compound over time.
There is a privacy cost. Every time you upload a file to an AI service, you are sending that document to the provider's servers. For a single file, you can make a considered decision about whether that is acceptable. When you are uploading the same sensitive document repeatedly across dozens of conversations, the surface area of exposure grows. The question of what data AI tools collect becomes more pressing when upload is the only mechanism available.
There is a format cost. Not every file type uploads cleanly. Spreadsheets lose their structure. Slide decks are flattened into text. Images in documents are sometimes ignored. Handwritten notes and voice memos cannot be uploaded at all in most interfaces. The upload slot favours a narrow band of file types, which means large parts of your working library are excluded from AI conversations entirely.
There is an attention cost. The minutes spent locating, selecting, and uploading files are minutes spent on logistics rather than on the question you came to ask. Over a working week, this adds up to a meaningful amount of time. Over a year, it represents a pattern of work that treats AI as a tool you feed rather than a tool that knows your context. Learning how to organise digital files can reduce some of this friction, but the root cause is architectural, not organisational.
How persistent AI access works
The alternative to uploading files for every conversation is to give AI standing access to your content. This does not mean sending your entire file library to a model. It means building a layer between your files and the model that handles retrieval on your behalf.
The technical foundation is retrieval-augmented generation, or RAG. In a RAG system, your files are indexed and stored separately from the language model. When you ask a question, the system searches your indexed content, identifies the passages most relevant to your query, and passes those passages to the model as context. The model reasons over the retrieved material and generates its response.
The important distinction is what happens at each step. Your files are not sitting inside the model. They are sitting in an index that the model can query. The index persists between conversations, so you never need to re-upload anything. New files are added to the index as they arrive, whether you save them manually or they sync from a connected app. And the retrieval step means the model sees only the relevant excerpts, not your entire library, which is both more efficient and more private.
This is how Fabric's AI assistant works. You connect the places where your files already live, including Google Drive, Dropbox, Notion, Gmail, and more than 50 other services, and Fabric indexes everything using semantic search. When you ask a question, the assistant retrieves the right material from across all your sources and passes it to the model. Files stay where they are. Nothing is duplicated or re-uploaded. The retrieval layer handles the work that you used to do by hand.
Because Fabric is a context layer rather than a single model, it works with AI from multiple providers: Claude, ChatGPT, Gemini, DeepSeek, and Grok. You choose the model suited to each task, and every one of them draws on the same indexed library. The context is yours; the model is whichever one fits.
Connections instead of uploads
The shift from uploading files to connecting sources changes the shape of what you can ask AI to do.
With uploads, you are limited to the files you brought into that specific conversation. With connections, your entire library is available for every question. You can ask something that spans a contract in Google Drive, meeting notes in Notion, and an email thread in Gmail, and the retrieval layer pulls the relevant pieces from each source without you needing to know which file holds the answer.
Fabric's connections handle the sync. When a new document appears in your connected Google Drive, it is indexed automatically. When a Notion page is updated, the index reflects the change. This means the AI's view of your content stays current without any manual intervention. You do not need to remember to re-upload an updated version of a file. The learn page on how to consolidate cloud storage covers the broader benefits of unifying scattered sources.
For tasks that go beyond retrieval, Fabric's agents can take actions across your connected tools. You grant an agent (Instinct, Muse, or Claude) scoped, time-limited access to a specific service, say reading a Slack channel or checking your Gmail, and it carries out the task within those boundaries. This is different from giving an AI blanket access to everything: you approve each scope and duration before the agent acts.
For files that do not live in a connected app, Fabric provides a cloud drive where you can store content directly. It requires zero local storage, streaming files on demand rather than syncing them to your device. Everything stored there is indexed and searchable in the same way as connected content. And quick capture lets you bring in web pages, screenshots, voice memos, and handwritten notes, all of which become part of the searchable index. The comparison between local and cloud storage is worth reading if you are weighing where your files should live.
MCP: portable context beyond one interface
One concern with any workspace-centric approach is lock-in. If your files are indexed in Fabric, does that mean you can only use Fabric's interface to access them?
This is where MCP, the Model Context Protocol, changes the picture. Fabric's MCP integration allows external AI tools to query your Fabric workspace directly. If you are using Claude in a desktop app, a coding assistant in your editor, or any other tool that supports MCP, it can request context from your Fabric library. The retrieval happens through Fabric's index, but the query can originate from anywhere.
This matters because AI is not a single interface. People use different tools for different tasks. A developer might ask questions about project documentation from within their code editor. A researcher might use a specialist tool for literature review. MCP means your indexed content is available to all of these tools, not just the one where you happen to have uploaded a file. Your context becomes portable.
The page on how to connect Google Drive to ChatGPT or Claude walks through how these integrations work in practice, and the guide on how to give ChatGPT memory of your notes covers a related use case.
Privacy when files stay in place
Repeated uploading is not just tedious; it is a privacy pattern worth examining. Every upload sends a copy of your file to the AI provider. If you upload the same document to ChatGPT ten times across ten conversations, that is ten transmissions of potentially sensitive content. You may trust the provider, but the exposure is cumulative.
A retrieval-based system works differently. Your files are stored in one place, either in Fabric's secure cloud drive or in the connected services where they already live. At query time, only the relevant excerpts are sent to the model provider, not the full file and not your entire library. The retrieval layer acts as a filter, minimising what the model sees.
For people working with confidential documents, client files, or anything covered by regulatory obligations, this architecture reduces the attack surface substantially. The guide on how to use AI on confidential documents safely goes deeper on the considerations involved. Fabric's smart organisation also helps by tagging and categorising content as it arrives, so you maintain visibility over what is in your workspace without needing to audit files manually.
Getting started
Moving away from the upload cycle does not require abandoning the AI tools you already use. It requires adding a layer underneath them.
Start by connecting your primary file sources to Fabric. If your work lives in Google Drive, connect Google Drive. If you use Notion for notes and Dropbox for project files, connect both. The sync is automatic and ongoing.
Once your sources are connected, ask the AI assistant a question that would normally require you to upload a file. You will find that the relevant content is already available. The retrieval layer handles the search, the model handles the reasoning, and you never open a file picker.
For files scattered across your devices, the cloud drive gives them a permanent, indexed home. For content you encounter on the web or capture on your phone, quick capture brings it into the same searchable space. And if you want your other AI tools to benefit from the same context, the MCP integration extends your indexed library to any compatible tool. Over time, Fabric's self-writing documentation can turn your accumulated content into maintained wikis, so the knowledge spread across files and notes becomes browsable without manual curation.
The upload button is not going away. But it should be the exception, not the workflow.
Frequently asked questions
Why do I have to re-upload the same file to ChatGPT every time?
ChatGPT does not retain files between conversations. Each session starts with an empty context, so the model has no memory of documents you shared previously. This is by design: the model's context window is session-scoped, not persistent. A workspace like Fabric maintains a persistent index of your files, so content is always available without re-uploading.
What are the file size and format limits for uploading to AI tools?
Limits vary by provider and plan. ChatGPT Plus supports individual files up to a certain size, with caps on the number of attachments per conversation. Claude has its own limits. In both cases, spreadsheets, slide decks, images, and audio files are handled inconsistently or not at all. A retrieval-based system like Fabric indexes files of many types and sizes, removing per-conversation constraints.
Is it safe to upload confidential documents to ChatGPT or Claude?
Safety depends on the provider's data policies and your plan tier. Enterprise plans typically offer stricter data handling than free or consumer tiers. However, each upload sends a copy to the provider's servers. A retrieval-based approach like Fabric's sends only relevant excerpts at query time, reducing exposure. The page on using AI on confidential documents safely covers this in depth.
What is retrieval-augmented generation and how does it replace uploads?
RAG is a technique where a search layer sits between your files and the language model. Instead of feeding the model an entire document, the system searches your indexed content for passages relevant to your query and passes only those passages to the model. This scales to large libraries, persists across sessions, and avoids the need to upload anything. The RAG explainer has the full details.
Can I still use ChatGPT or Claude if my files are in Fabric?
Yes. Fabric is not a replacement for ChatGPT or Claude. It is the context layer that sits underneath them. Fabric's AI assistant gives you access to models from OpenAI, Anthropic, Google, and others, all with access to your connected files. You can also use Fabric's MCP integration to bring your indexed content into other AI tools directly.
How does Fabric connect to my existing cloud storage?
Fabric's connections feature integrates with Google Drive, Dropbox, Notion, Gmail, Slack, and more than 50 other apps. Files stay in their original location. Fabric indexes them for semantic search and AI retrieval without moving or duplicating them. New and updated files are synced automatically.
What is MCP and why does it matter for this workflow?
MCP, the Model Context Protocol, is a standard that lets AI tools request context from external sources. Fabric's MCP integration means that any MCP-compatible tool, whether it is a desktop AI assistant, a coding environment, or another application, can query your Fabric workspace for relevant content. This makes your indexed files available across tools, not just within Fabric's own interface.
Does Fabric store my files or just index them?
Both, depending on the source. Files from connected apps like Google Drive remain in their original location; Fabric indexes them without creating copies. Files you upload directly or capture with quick capture are stored in Fabric's cloud drive. In both cases, content is indexed for semantic search and AI retrieval.
What types of content can Fabric index that I cannot upload to ChatGPT?
Fabric indexes PDFs, documents, spreadsheets, presentations, images, web clips, voice memos, handwritten notes, and content from connected apps like Slack and Gmail. Many of these formats are difficult or impossible to upload to ChatGPT or Claude directly. The semantic index treats all of them as searchable, queryable content.
Will this workflow become unnecessary as AI context windows get larger?
Larger context windows reduce some upload constraints, but they do not solve the core problem. Even a model that can hold a million tokens in its context window still requires you to select and upload the right files for each conversation. It still has no persistent memory of your content between sessions. And sending large volumes of data to a model for every question is less efficient and less private than retrieving only the relevant excerpts. The retrieval-based approach remains valuable regardless of context window size.
Related pages
You might be interested in:
How to let clients proof and choose photos online
How to search images by colour
How to send a different tracking link to each recipient
How to stop search engines indexing a shared file
How to share files on your own domain
How to turn a folder into a website
How to edit video straight from a cloud drive
How to access cloud files without using disk space