Your AI Agent can only answer well if it has the right information to draw on. The knowledge base is where you give it that: your FAQs, product details, pricing, policies, and anything else a follower might ask about.
To get there, click AI Agent in the sidebar. You'll find two cards under the Knowledge heading: Knowledge access and behaviour and Training and knowledge.
Knowledge access and behaviour
This setting controls what the agent does with a question your uploaded knowledge doesn't cover. There are two choices:
Always have an answer. The agent uses your uploaded knowledge plus the model's own general knowledge. If a question isn't in your documents, it still tries to answer. Good for accounts that field a wide range of conversational questions, but it can produce replies that drift off-topic.
Limit to instructions. The agent sticks to what you've taught it and your stated goals. If it can't find a relevant answer, it says so rather than guessing. This is the default, and the safer choice for most businesses.
When you choose Limit to instructions, a sub-option appears: If the agent cannot reply, trigger this automation. Pick a scenario here and the agent runs it whenever it can't answer, so you can route the contact to a human, send a holding message, or tag them for follow-up.
Training and knowledge
This is where you add the content the agent learns from. Each piece is called a knowledge item.
Adding a knowledge item
Click Import new item. A window opens with three source types:
Manual text. Paste content straight into a text field (up to 5,000 characters). Good for FAQs, quick facts, or anything you'd rather write by hand.
Text document. Upload a file. The accepted formats are .txt, .pdf, .doc, and .docx. Use this for existing materials like product guides, price lists, or support scripts.
Website. Enter a URL and Inrō imports the content from that page. Good for your pricing page, About page, or public help docs.
After you add the content, click Save item. It appears in your knowledge grid with a status of Processing, then Ready once Inrō has indexed it.
⚠️ A website item imports a single page, not your whole site. To cover several pages, add a separate item for each URL.
What the agent does with it
Inrō breaks each item into smaller chunks and indexes them. When a message comes in, the agent searches across all your items for the most relevant pieces and uses those to write its reply. You don't need to organise or tag anything.
When an item fails
If an item shows Failed, the content couldn't be read. The usual causes are an unsupported file type, a PDF or document with no readable text (for example, a scan that's really an image), or a web page that couldn't be reached or had no text to pull. Delete the item and try again, for example by pasting the text in manually instead.
Editing and removing items
Click any item in the grid to edit or delete it. Changes take effect right away for all future conversations.
What to add to your knowledge base
A strong starting set usually includes:
FAQs: the 10 to 20 questions you get most, with your standard answers.
Product or service details: what you offer, how it works, what's included.
Pricing: your prices or packages, or where to find them if they're not public.
Policies: refunds, shipping, turnaround times, booking terms.
Contact and booking info: how to reach you, how to book.
Social proof: key results or testimonials in plain text.
You don't need to add everything at once. Start with your most common questions and fill gaps as you spot them.
🐾 Netsuke's Tips
Write manual text entries the way you'd answer out loud, not as a formal document. The agent mirrors the tone of what it reads, so conversational source material produces conversational replies.
If you have a detailed FAQ page, import it with the Website option first. It's the quickest way to a solid base, and you can refine individual answers later with manual text.
After adding new knowledge, use Chat with Agent to ask the exact questions the new content should answer. It's the fastest way to confirm the import worked.
If the agent keeps getting one topic wrong, add a manual text entry that addresses it directly. Specific entries tend to win over general ones when the agent searches for an answer.
There's also an option to include a text item in every reply rather than only when it's relevant. It needs your own OpenAI key and can slow replies down, so use it only for short, must-always-apply rules. See AI Agent Best Practices.
What's next?
With knowledge in place, set up actions so the agent can do things mid-conversation, not only answer. Go to Intent-Based Actions: Links, Folders & Scenario Triggers to configure what happens when the agent detects a specific intent.