Quick Start

Knowledge

Feed your agents your own content so answers come from your business, not guesswork.

4 min readSidebar: KnowledgeSwitch to Complete Manual
Who can see this: Hidden from agents and supervisors.

Knowledge is what makes your AI answer from your business instead of guessing. Every chat widget and call agent can be attached to a knowledge base, and answers come back with inline citations. The sidebar section is Knowledge, and its pages are Overview, Documents, Sources, Web Crawler, and Settings.

This is the highest-leverage section in the product. An hour spent here beats a day spent editing prompts. It is hidden from agents and supervisors.

Step 1 — Create a knowledge base#

Press Add Knowledge Base in the top right, give it a Name and a Description, then Create.

1
Step 1. Knowledge → Overview. One card per knowledge base, with its embedding state.

Most workspaces need only one. Create several when you have genuinely separate domains — different brands, or a public FAQ versus internal procedures — because a widget attaches to exactly one.

The Overview tiles summarise everything: Total Knowledge Bases, Active Sources, Total Documents, Last Sync, and Embedding Status.

Step 2 — Crawl your website#

Go to Knowledge → Web Crawler. Under Start New Crawl, enter a Website URL (example.com or the full address), pick your Knowledge Base, and press Start Crawl.

2
Step 2. Web Crawler. Jobs report live progress as pages are processed.

Show Settings exposes the controls, which are already set sensibly:

SettingDefaultRange
Max Depth51–10
Max Pages1001–1000
Delay (ms)10000–10000
Chunk Size1000100–5000
Respect robots.txton
Follow external linksoff

Jobs appear below with live progress — pages found, pages processed, and a percentage — and can be filtered by running, pending, completed, failed, or cancelled. A running job can be Stopped; a finished one can be deleted.

One important limitation: the crawler skips PDFs, Word files, and other documents linked from your pages. It reads HTML. Anything in a file has to be uploaded in the next step.

Step 3 — Upload the documents that are not on your website#

Knowledge → Documents. Pick your knowledge base on the left, then Upload File.

3
Step 3. Documents. Knowledge bases, their documents, and the selected document's content.

Supported files are PDF, Word (.docx), and plain-text formats such as .txt, .md, and .csv.

Anything else is read as raw text, which usually produces gibberish or an empty document without reporting an error — so check the document after uploading. Legacy .doc files are the common casualty; resave them as .docx.

This is usually where the answers your customers actually want live: price lists, policies, warranty terms, delivery rules. Add Manually is there for writing an answer directly, which is the fastest fix when you spot a gap.

Step 4 — Generate embeddings#

Back on Knowledge → Overview, each card shows Embeddings: followed by its state. Press the button beside it.

StateButtonMeaning
No DocumentsGenerate EmbeddingsNothing to embed yet
Not GeneratedGenerate EmbeddingsContent exists but has never been embedded
Needs UpdateUpdate EmbeddingsDocuments changed since the last run
CompleteRegenerate EmbeddingsUp to date

Content is not searchable until this finishes. A knowledge base full of documents with no embeddings answers nothing — this is the single most common reason a widget "ignores" content that is plainly there.

Watch for Needs Update afterwards. Every time you add or edit a document, the card flips to that state and you need to run it again.

Step 5 — Attach it, then test it#

Attach the knowledge base to a widget when you create or edit it — see Chat for chat widgets and AI Calls for voice.

Then ask your widget a question only your documents can answer.

If the answer is vague, add content — do not rewrite the prompt. A missing fact cannot be prompted into existence, and the fastest fix is almost always a short manual document that states the answer plainly.

Keeping it current#

Two mechanisms, and they are different:

Sources (Knowledge → Sources) are recurring feeds. Add Source takes a Type (Website, Sitemap, RSS Feed, API Endpoint), a Name, a URL, and a Sync Frequency of Manual, Daily, Weekly, or Monthly. Use these for content that changes.

Settings (Knowledge → Settings) holds per-knowledge-base retrieval options: Chunk Size, Chunk Overlap, Language, and Auto-update sources and embeddings. Leave Auto-update on and the maintenance mostly takes care of itself.

Where to go next#

If you sell on Shopify, each knowledge base card has a Shopify catalogue toggle. Switch it on and your products are embedded too, so the AI can answer about stock and specifications. It needs a connected Shopify store — see Integrations.