Taskade: How to Train an AI Agent on Your Docs

The word "training" is misleading here, and the confusion causes most of the disappointment. You are not retraining a model on your company. You are giving an agent a searchable library it can consult mid-answer.

That distinction changes what you should upload, how you should structure it, and why a tidy ten-page document often beats a two-hundred-page one. This guide covers the mechanics first, then the part that decides whether it works.

Taskade describes the mechanism openly: documents are split into chunks, each chunk is converted into a numeric representation, and at question time the most relevant chunks are pulled into the model's context before it answers. The industry name for this is retrieval-augmented generation.

What can you feed a Taskade agent?

Taskade documents five source types on the Knowledge tab, and they behave differently enough that the choice matters.

• Files you upload. Their documentation lists support for documents such as PDF, DOCX, TXT, MD, and RTF, spreadsheets such as XLSX, CSV, and TSV, and PPTX presentations, with a per-file size cap noted at 5 MB at the time of checking.
• Projects from your workspace. Live Taskade projects, read as structured content. Useful because the agent sees your current list rather than a stale export.
• External web links. Pages the agent should be able to quote from, such as your public documentation or a policy page.
• YouTube videos. Added by URL and transcribed automatically, which is the fastest way to get a recorded walkthrough into an agent.
• Cloud storage. Google Drive, Dropbox, and Box connections, plus files already sitting in the Media Manager.

Their own guidance ranks these by usefulness rather than treating them as equals. Custom instructions are always in play, workspace projects give precise line-level detail at a higher token cost, uploaded documents supply broad background, and web or video sources fill in the edges.

Add knowledge to an agent

1. Open the agent and go to the Knowledge tab — Knowledge is a tab in the agent editor, alongside General, Tools, and Commands.
2. Turn the Knowledge toggle on — Nothing you add takes effect while the toggle is off. This catches people out more often than it should.
3. Add your sources — Drag and drop files, pull items from the Media Manager, add workspace projects, paste external links, or add a YouTube URL. Connect Google Drive, Dropbox, or Box if the files live there.
4. Save, then ask the agent a question you know the answer to — Pick something specific from page four of a document, not something general. A general question can be answered from the model's own knowledge and tells you nothing.
5. Run a real test set before you rely on it — Taskade's own best-practice guidance suggests running at least 10 to 15 test queries across the different categories of question you expect before you put the agent in front of anyone.
6. Schedule the refresh — Use the Add Knowledge to Agent automation action so the library updates itself rather than depending on you remembering.

Here is the counterintuitive part, and it is the single highest-value thing in this article. Taskade's own guidance states that a focused set of five to ten well-organised documents often outperforms fifty loosely related PDFs.

Why does document structure change the answers?

Because the agent never reads your document end to end. It retrieves fragments, and a fragment is only useful if it makes sense on its own.

A heading that says "Refund window: 30 days" followed by three sentences of explanation survives being cut out and read alone. A sentence that says "in that case the above applies" does not, because "the above" was in a different fragment.

So the practical rules follow from the mechanism. Use clear headings, keep paragraphs short, split files by topic rather than by department, and repeat the key noun instead of writing "it" or "that policy".

How do you keep the knowledge current?

Manually re-uploading files is where most knowledge bases quietly rot. Taskade documents an automation action called Add Knowledge to Agent that removes the manual step.

You add it as a step in an automation, choose the Agent it should feed, pick the Knowledge Type from website link, YouTube, file, or plaintext, supply the source input, and optionally give the entry a Title. Their documentation pairs it with a schedule trigger for recurring refreshes, or with a new-file trigger from Google Drive so a dropped document trains the agent by itself.

If curating a knowledge base is not what you wanted

Splitting documents by topic, rewriting headings so fragments stand alone, running fifteen test queries, and scheduling a refresh is a real project. It is the right project when you want a specialist that answers strictly from a library you control, and Taskade gives you honest, well-documented tools for it.

Work like that is what we removed at Sistava. You hire an AI Employee and tell it about your business the way you would brief a new starter, in plain English. It asks you when it is unsure rather than retrieving a wrong fragment silently, and there is no library to restructure.

Neither approach is universally better. Pick the library if precision over a fixed corpus is the point, and pick the briefing if the corpus is really just "how we do things here".

Sources

• Taskade Help Center, Custom AI Agents, including the Knowledge tab
• Taskade blog, Train Custom AI Agents: files, data, web links and automation
• Taskade blog, Best Practices for Training AI Agents With Knowledge in Taskade
• Taskade Help Center, Add Knowledge to Agent (Automation Action)
• Taskade pricing page, plan limits as listed on 17 August 2026

Curate the library if the precision is the point. If what you wanted was somebody who already understands the business and gets on with it, that is a different job description.