Add content to your knowledge base

The four ways to load help articles, files, web pages and FAQs into the AI, plus the per-plan limits.

The knowledge base is the set of facts your Assist AI answers from. Every account has its own, and content you add is scoped to your account — it is never shared with other tenants. There is no separate "on" switch: as soon as you add an item, the AI can use it to answer questions on your widget and connected channels.

Open the Knowledge Base section of the Assist dashboard. It has two tabs:

Documents — free-form articles, policies, uploaded files, and crawled web pages.
FAQs — short question-and-answer pairs.

Four ways to add content

1. Paste text

Give the item a title (for example, "Refund Policy") and paste the body. This is the best fit for policies, product details, onboarding steps — anything you would tell a customer in your own words.

2. Upload files

Drag and drop onto the upload area, or click to browse. Supported file types are .txt, .md, .json, .csv, and .docx. Word documents (.docx) are parsed to plain text automatically; the other formats are read as text. Drop a single file and it fills the form so you can edit the title and content before saving. Drop several at once and each is uploaded as its own separate document.

There is no PDF support — .pdf is not an accepted type, and a PDF read as raw text comes through as garbage. Convert the PDF to a .docx, or copy the text and paste it in as a text item.

3. Crawl a web page

Switch to the website mode and enter a URL such as https://docs.yourcompany.com (if you leave off https:// it is added for you). Wallu fetches that one page, strips out the navigation, header, footer, forms, scripts and styles, and saves the readable text as a document.

Two things to know before you rely on it:

It is a single-page fetch, not a whole-site crawler. It does not follow links or pull in your entire site — add each page you want, one at a time.
It times out after 55 seconds (with a 20-second limit on fetching the page itself). Sites that block bots or respond slowly return a timeout error; paste the content manually in that case.

A crawled page is capped at 20,000 characters — much smaller than the per-item limit further down. A long docs page is trimmed to roughly its first 20,000 characters, so if the page you want is longer than that, split it across several fetches or paste the text in yourself.

4. Add FAQs

On the FAQs tab you can add a question and answer one at a time, or bulk-paste several pairs at once in this format:

Q: How do I reset my password?

A: Go to settings and click Reset Password.

Q: What is your refund policy?

A: We offer a 30-day money-back guarantee.

Each Q: / A: pair becomes its own FAQ entry.

Editing and feedback

Every item can be edited or deleted from the list. Long documents open in an in-place editor with search-in-content highlighting so you can jump to the part you want to change.

When a visitor rates an answer with thumbs up or down in the chat widget, that feedback is tied back to the knowledge-base item that produced the answer. So each item shows how many times it was marked helpful versus not helpful — a quick way to spot weak or outdated articles and fix them.

Limits

Content is capped per Assist plan. The caps are generous — most accounts never reach them — but they exist to keep storage and embedding cost bounded.

Per item: 200,000 characters (roughly 50,000 tokens). Split anything larger into separate articles.
Free: 15 items, 300,000 characters total.
Lite: 50 items, 1,000,000 characters total.
Starter: 200 items, 5,000,000 characters total.
Pro: 1,000 items, 20,000,000 characters total.
Enterprise: 100,000 items, 500,000,000 characters total.

Sub-accounts inherit the owner account's plan caps. When you hit a limit, adding is blocked with a message telling you to delete items or upgrade — nothing you have already stored is touched.

With content in place, see getting started to get the widget live, or read the next article, How the AI finds and uses your content, to understand what happens at answer time.

How the AI finds and uses your content

Embeddings, semantic matching, and how retrieved knowledge feeds the AI's answers and its confidence.

You never manage an index by hand. When you add or edit an item, Wallu automatically turns its text into a semantic vector (an embedding) and stores it with the item. That vector is what lets the AI match a customer's question to the right content even when the wording is completely different.

Embeddings

Each item is embedded with Google's gemini-embedding-001 model into a 3,072-dimension vector.
Long items are handled without dropping anything: the text is split into roughly 6,000-character chunks (up to 16 of them), each chunk is embedded, and the vectors are averaged into one. So a long policy is represented by its whole content, not just the opening paragraph.
Re-embedding happens automatically whenever you change an item's content — there is no rebuild step to trigger.

How a question gets answered

On every incoming message, before the AI writes its reply:

The customer's question is embedded the same way.
That vector is compared against your stored item vectors by cosine similarity — closeness in meaning, not keyword overlap.
The most relevant items are pulled and injected into the AI's prompt inside a clearly marked === KNOWLEDGE BASE === block, with an explicit instruction to answer from that content and not make up facts.

The retrieved text is trimmed to a token budget before it is sent, so a handful of the closest items are used per reply rather than your entire knowledge base. In the chat-widget path, up to the top 5 matching documents are used, each trimmed to about 5,000 characters (15,000 total). If your knowledge base is small enough to fit the budget, the AI is simply given all of it; once it grows past that, retrieval sends only the items most relevant to each specific question.

What this means in practice

Write focused items. One clear topic per document retrieves better than a giant catch-all page, because it is the closest-matching content that surfaces.
Mirror the phrasing customers actually use. Semantic matching is forgiving, but an FAQ written like a real question ("Can I use this on a Mac?") anchors the answer strongly.
Avoid contradictions. If two items disagree, the AI may surface either one. Keep a single source of truth and delete stale versions.
New content is live immediately. The item is embedded on save and available to the very next message — no publishing or reindex step.

Retrieval and confidence

The similarity score of the best-matching item feeds Assist's confidence logic. A strong match produces a high-confidence, grounded answer. When nothing in your knowledge base is close to the question, confidence drops below the threshold — but the AI does not hand off on its own. Instead it still answers from what it has and appends a disclaimer inviting the visitor to type human to speak with an agent. The actual handoff fires only when the visitor asks for one: escalation is triggered by the customer's message matching human-request keywords (words like "human", "agent", or "escalate"). If that user-requested handoff happens while confidence was low, it is logged with the reason "AI was unsure of the answer."

That is why coverage matters: the more of your real content lives here, the more questions get confident, sourced answers instead of a low-confidence reply that nudges the visitor toward a human. For where those answers appear and how escalation works, see channels and tickets. To review handoffs and conversations, use the unified inbox.