Skip to main content
The knowledge base is the system that makes your assistants smart about your specific business. Instead of relying only on general AI knowledge, your assistants can read live data and search through your uploaded documents and curated Q&A pairs to find accurate, relevant information before responding. The Knowledge Base page is organized into two sections, each shown as a grid of cards — one card per source. Every card shows a quick summary (a connection or sync status, plus a count of what’s in it), and clicking a card opens that source’s own management page where you add and manage its content.
  • Realtime data — Connect live sources like Shopify, WooCommerce, and Spreadsheets. These are checked live every time someone chats (never copied into your knowledge base), so answers are always current.
  • Static knowledge — Content you add once, saved so your assistant can search it. Each source is its own card:
    • Files — Upload documents (PDF, Word, Excel, PowerPoint, text, Markdown, CSV) your assistant can search by meaning, not just keywords
    • Websites — Crawl and sync pages from your site
    • Text — Add freeform snippets directly
    • Q&A Pairs — Create structured question-and-answer entries for precise, high-confidence responses
Each static card shows how many items it holds (e.g. “3 files”) and a sync status badge, with an Add button to start adding content. Opening a card takes you to a focused page for just that source — for example, the Files card opens a page where you upload and manage only your files.

​
Realtime Data Sources

Realtime data sources connect your assistant to systems that change constantly — your store catalog, orders, and fulfillment status. Unlike static knowledge, this data is never copied into your knowledge base; it’s checked live every time someone chats, so customers always get the most current answer. Each realtime source card shows a connection-health badge so you can see at a glance whether it’s connected, and a Test connection button to verify it on demand.
Shopify, WooCommerce, and Spreadsheets used to live under Channels. They now live here, under Knowledge Base → Realtime data. Older links to the channel pages redirect here automatically. Availability depends on your plan.

​
How It Works

IllumiChat uses Retrieval-Augmented Generation (RAG) to connect your documents to your assistants:
1

Add your content

Open your assistant’s Knowledge Base, then under Static knowledge open the Files card and drop your documents onto the upload area. See Supported File Types for the full list.
2

Automatic processing

IllumiChat extracts the text from your documents and splits it into smaller, meaningful chunks.
3

Embedding and indexing

Each chunk is converted into a vector embedding — a numerical representation that captures the meaning of the text. These are stored in a vector database for fast retrieval.
4

User asks a question

When someone sends a message to your assistant, IllumiChat searches the vector database for document chunks relevant to the question.
5

Context-aware response

The assistant receives the relevant chunks as context alongside the user’s question, then generates an accurate answer grounded in your actual documents.

​
Fallback Response

Your assistant answers only from sources you control — your knowledge base, plus any connected store or live data source. It never falls back on general AI knowledge to fill a gap. The fallback response is the message it sends when none of those sources can answer the question. You’ll find it on your assistant’s Knowledge Base page, under Knowledge settings → Fallback response.
Leaving the field blank is perfectly fine. Your assistant then sends the default: “I don’t have information about that in my knowledge base. Please contact our support team for assistance.” Set your own when you want to point customers somewhere specific.

​
When It’s Sent

The fallback is sent word for word. Your assistant won’t paraphrase it, soften it, prefix an apology, or slip in a guess of its own.
The fallback response is not an error message. A temporary outage produces a different reply on purpose, so “we don’t have that information” never gets confused with “something is broken.” Avoid copy like “Sorry, something went wrong” — it would be shown for perfectly healthy questions that simply fall outside your content.
Because it’s sent verbatim, treat it as a piece of customer-facing copy in your brand’s voice, not as a system string.

​
Examples

I don’t have that one in my notes yet. Our support team can help — email support@yourstore.com or reply here and we’ll pass it along.
Best when you have a staffed inbox. It sets an expectation without promising a response time you can’t keep.
I can’t answer that one, but I can help with orders, shipping, returns, and product questions. Want me to look up an order?
Best for a widget on a storefront. Naming your assistant’s strengths turns a dead end into a second attempt, and cuts repeat “can you help with X?” messages.
That’s outside what I have documented. Our docs cover it in more depth at https://yourcompany.com/docs, or I can connect you with our team.
Best for technical products where the real answer is long. Use a bare URL rather than a Markdown link — plain-text channels like SMS and WhatsApp show whatever you type literally.
I’m not able to answer that one. Let me get a teammate — type agent and I’ll bring someone in.
Best when Live Chat is enabled and someone is actually available. See Universal Inbox for staffing the handoff.
That isn’t in the handbook I have access to. Ask in #people-ops, or file a request if it’s something we should document.
Best for internal assistants, where the useful move is naming the human channel instead of apologizing.

​
Writing a Good Fallback

A good fallback does three things: admits the gap plainly, gives one concrete next step, and sounds like your brand.
  • Give exactly one next step. An email address, a URL, or a keyword like “agent.” Two or three options make people pick instead of act.
  • Say what you can do. “I can help with orders, shipping, and returns” recovers the conversation. “I don’t know” ends it.
  • Skip the hedging. It’s sent verbatim, so “Unfortunately, I’m terribly sorry, but I’m afraid…” lands as one long apology every time.
  • Don’t blame the customer. “Try rephrasing your question” implies they asked it wrong; usually the content is just missing.
  • Watch what triggers it. A fallback firing often is a content gap, not a copy problem. Add a Q&A pair or a document covering the question.

​
Limits

  • Up to 500 characters
  • One message per assistant, shared across every channel — widget, SMS, WhatsApp, Messenger, Instagram, and email
  • Applies once you have a knowledge base; an assistant with no knowledge base has nothing to fall back from

​
Supported File Types

Files can be up to 50 MB each. Total storage across all your knowledge sources depends on your plan — the Files page shows a used-of-available meter.

​
Formats that aren’t supported

If you upload something unsupported, the document shows Needs attention with an explanation rather than failing silently.

​
Uploading Documents

1

Open your assistant's Knowledge Base

Open the assistant you want to teach, then go to its Knowledge Base.
2

Open the Files card

Under Static knowledge, click the Files card (or its Add button) to open the file management page.
3

Add your files

Drag files straight onto the upload area, or click it to browse your computer. You can also pull files from Google Drive, Dropbox, OneDrive, Box, or a direct URL using the buttons beside the drop area. Multiple files at once is fine.
4

Monitor status

Each document shows a status so you know where it is. Once it reads In use, the assistant can search it automatically — there’s no separate linking step.

​
Uploading from Google Drive and other cloud storage

Google Workspace files (Docs, Sheets, Slides) aren’t ordinary files — they live in Google’s format and have to be converted on the way in. IllumiChat picks the conversion that preserves the most text: Files already stored in a normal format (a PDF sitting in your Drive, a .docx in Dropbox) are taken exactly as they are, with no conversion.

​
Extraction Caveats

Your assistant searches the text of a document. What that text looks like after extraction varies by format, and a few cases are worth knowing about before you rely on them.
If an answer looks wrong or incomplete, the extracted text is the first thing to suspect. Uploading the same content in a cleaner format usually fixes it outright.
Spreadsheets lose their shape. Cells come through as a flat run of text, so the assistant reads the values but not the grid. Formulas contribute their last saved result, and charts, pivot tables and images contribute nothing. A spreadsheet of policies or FAQs works well; one that answers questions only through its layout does not. For order status or inventory that changes, connect a Spreadsheet under Realtime data instead of uploading a copy. Presentations favor words over pictures. Slide text and speaker notes are read; diagrams, screenshots and charts aren’t. A deck that carries its meaning visually will contribute far less than its page count suggests. PDFs vary the most. A PDF exported from a word processor extracts cleanly. Multi-column layouts, tables, sidebars and footnotes can extract in a jumbled reading order, because a PDF records where text sits on the page rather than how it should be read. Where you have the original .docx, .xlsx or .pptx, upload that instead — it extracts more reliably than a PDF made from it. Scanned pages contain no text. A scan or phone photo saved as a PDF is a picture of a document. There is no OCR, so nothing is extracted and the document ends as Needs attention. A few Google file types have no text form. Drawings and Jamboard boards can only be converted to PDF, and since they’re visual, the result is usually near-empty. Copy anything important into a Doc first. HTML brings the page furniture with it. Tags are stripped, but navigation menus, cookie banners and footers become part of the text. For website content, use the Websites source, which is built for it, rather than uploading saved pages. Very large files are refused, not truncated. Anything over 50 MB is rejected at upload. Split it into sections and upload them separately — smaller documents also retrieve more precisely.

​
Document Status

Every item in your static knowledge shows one of four statuses. The same four words are used on the source cards, in the document lists, and in the status filter.
Sync time depends on document size and complexity. Most documents reach In use within a few minutes.

​
Keeping Static Knowledge Fresh

Static sources (files, websites, text, and Q&A) are synced once when you add them. To keep them current as your underlying content changes, each item shows a sync status badge and offers two ways to refresh it:
This badge reports the last sync attempt. The status above (In use / Needs attention) reports whether your assistant can answer from the item at all. A document can be In use while serving its previous content and still show Sync failed for its most recent update.

​
Re-sync now

Use the Re-sync now button on any item to re-sync it immediately — for example, after you’ve updated a webpage or replaced a document’s contents.

​
Auto-sync

Set an auto-sync cadence per item so IllumiChat re-syncs it on a schedule:
Auto-sync is a paid feature. It’s off on the Free plan and available on Pro and Enterprise. If you’ve just upgraded or a new release added the feature, your workspace admin may need to refresh plan features before the cadence selector appears.

​
What Happens When You Delete an Assistant

When you delete an assistant, IllumiChat also cleans up the resources it owned so nothing is left behind:
  • Its knowledge base is removed — uploaded files and synced content are deleted, and the search index is reconciled automatically shortly afterward.
  • If it was connected to Shopify or WooCommerce, IllumiChat removes the widget it installed on your store and unregisters the webhooks it created there.
Cleanup runs on a best-effort basis: if one step can’t complete (for example, your store is temporarily unreachable), the assistant is still deleted and the rest of the cleanup proceeds. Messaging-channel connections (Messenger, Instagram, WhatsApp) are not yet detached automatically on deletion.

​
Best Practices

Documents that cover one topic thoroughly work better than large documents covering many topics. This helps the retrieval system find the most relevant chunks.
Well-structured documents with clear headings, bullet points, and short paragraphs produce better search results.
Keep your knowledge base current by replacing outdated documents. Stale information leads to inaccurate answers.
After uploading documents, test your assistant with the kinds of questions your users will actually ask.
Knowledge belongs to the assistant, so give each audience its own assistant instead of pooling everything into one. Keep customer-facing product docs on your support assistant, and internal HR policy on a separate private assistant.

​
Q&A Pairs

Q&A pairs let you curate precise answers for frequently asked questions. When a user’s question closely matches a stored Q&A pair, the assistant returns the curated answer directly — providing more concise, accurate responses than general document retrieval.

​
How Q&A Matching Works

IllumiChat uses a hybrid matching strategy for Q&A pairs:
  1. Direct match — Your question text is embedded and compared semantically against stored Q&A pairs. If the similarity score meets the confidence threshold, the paired answer is returned verbatim.
  2. RAG fallback — If there is no high-confidence direct match, Q&A content participates in standard document retrieval alongside your uploaded files, so partial or indirect matches can still surface.
This means Q&A pairs give you the best of both worlds: precise answers for known questions and fuzzy relevance for everything else.
Q&A pairs share the same document quota as your uploaded files. Each Q&A pair counts as one knowledge document.

​
Adding Q&A Pairs

1

Open your assistant's Knowledge Base

Open the assistant you want to teach, then go to its Knowledge Base.
2

Open the Q&A card

Under Static knowledge, click the Q&A Pairs card. It sits alongside the Files, Websites, and Text cards.
3

Create a new pair

Click Add Q&A and fill in the Question and Answer fields.
4

Save

Click Save. The Q&A pair is synced automatically and available to this assistant.

​
Managing Q&A Pairs

From the Q&A Pairs card, you can:
  • Edit an existing pair to update the question or answer
  • Delete pairs that are no longer needed
  • View all pairs in a list with the question and a preview of the answer
Start with Q&A pairs for your top 10-20 most common support questions. These are the questions your team answers repeatedly — giving your assistant curated answers for these dramatically improves response quality.

​
When to Use Q&A vs. Documents

​
Best Practices for Q&A Pairs

Use natural, conversational phrasing. If your users ask “How do I reset my password?” rather than “Password reset procedure,” write the question that way. The semantic matching works best when your stored question mirrors real user queries.
Write answers that fully address the question without unnecessary detail. If the answer requires a longer explanation, consider linking to a document or guide.
If a critical question can be asked in very different ways, consider adding separate Q&A pairs for each variation to maximize match accuracy.
As your product or policies change, update your Q&A pairs to stay current. Outdated answers erode user trust.
For the best support bot experience, combine Q&A pairs for your most common questions with uploaded FAQs, help center articles, and troubleshooting guides. This gives your assistant precise answers for known questions and broad knowledge for everything else.