Knowledge

Complete reference for the Knowledge section — knowledge bases, documents, sources, the web crawler, and retrieval settings.

Hidden from agents and supervisors.

For the fast path, use the Quick Start.

Section map#

/knowledge-base redirects to /knowledge-base/overview. Add Knowledge Base is available from every tab.

Roles#

Two things follow from this. Editors can build but not delete — a useful split, since deleting a knowledge base also removes every document and source under it. And the section is hidden from the sidebar for agents and supervisors rather than blocked at the route, so someone who knows the URL can still open the page; the mutations themselves are enforced on the server, which is where it matters.

How retrieval works#

Content goes through four stages, and knowing them explains almost every problem in this section.

The critical implication: a document is invisible until it has been embedded. Ingestion and embedding are separate, and the second is manual unless Auto-update is on. This is the cause of most "the AI ignores my content" reports.

Chunk size is a genuine trade-off. Small chunks retrieve precisely but can lose the context around a fact; large chunks keep context but dilute the match. The defaults are reasonable — change them only when you have a specific failure to fix.

One knowledge base or several

A widget or call agent attaches to exactly one knowledge base. Split only when the content should not mix — separate brands, or public versus internal material. Splitting content that belongs together means one widget can no longer answer questions that span both halves.

Overview#

Enlarge

Tiles

Creating and editing

Add Knowledge Base opens Create Knowledge Base with a Name (required) and a Description.

Edit Knowledge Base adds one control that is easy to miss:

Website crawling — Allow website sources to be crawled and embedded. When off, sync skips website sources.

With it off, Sync is disabled on the card and the tooltip explains why: Website crawling is disabled for this knowledge base. Enable it in Edit to sync website sources. Uploaded documents are unaffected.

Deleting warns: Are you sure you want to delete this knowledge base? This action cannot be undone. All associated documents and sources will also be removed.

Filters

A search box (Search knowledge bases...), a status filter (all, idle, syncing, completed, error), and Refresh. Empty: No knowledge bases found — Create your first knowledge base to get started.

While anything is syncing the page refreshes itself every 10 seconds, so you can watch progress without reloading.

Reading a card

Each card shows the name, a status pill, document and source counts, Embeddings: with its state, the Shopify catalogue row, and when it was last updated.

Embedding status

Needs Update is highlighted in amber because it is the state that quietly breaks answers. It appears the moment any document's created or updated timestamp is newer than the last embedding run — so it will appear after every edit, upload, and completed crawl.

On success: Embeddings generated successfully! {n} documents processed in {n}s.

Set Auto-update sources and embeddings in Settings to have this happen without you.

Card actions

The Shopify catalogue

Each card has a Shopify catalogue toggle — Embed the Shopify product catalogue into this knowledge base. Requires a connected Shopify store.

The row reads off, syncing…, error, or a product count. Switching it on starts embedding immediately (Product catalogue enabled — embedding started), and a Sync Catalogue button appears for manual refreshes.

This is what lets the AI answer product questions — specifications, variants, availability — during a chat or a call, without you writing any of it down.

If the FloGPT Shopify app has been uninstalled, the knowledge base moves to a deactivated state and every sync control is locked with the explanation: The floGPT Shopify app has been uninstalled from this store. Reinstall it to re-enable syncing.

See Integrations and FloGPT for Shopify.

Documents#

Enlarge

Three panes: knowledge bases, their documents, and the selected document.

Adding content

Supported uploads are PDF, Word .docx, and any plain-text format (.txt, .md, .csv and similar). Legacy .doc is not supported — resave as .docx.

There is no type filter on the picker and no enforced size cap, and — importantly — an unsupported file does not fail loudly. Anything the parser does not recognise is read as raw UTF-8, so a binary format lands as gibberish or as an empty document while the upload still reports success. Open the document after uploading an unusual format and confirm the Content looks like prose. Gibberish that reaches the embedding stage degrades retrieval across the whole knowledge base, so delete it rather than leaving it in place.

Add New Document takes:

Writing a short manual document is the fastest way to fix a wrong or missing answer — much faster than editing a source page and re-crawling.

Browsing

A search box (Search documents...) matches title, summary, and tags, and a category filter offers All Categories plus whatever categories exist on your documents.

The list shows 5 documents at a time with a Load more ({n} remaining) control, ending in All {n} documents. With a large knowledge base, search rather than scroll.

Each row shows the title, the summary (or No summary), a word count, its category, and a status pill.

Viewing and editing

The detail pane shows the title, status, Type, Summary, Content, and a metadata row of Language, Word Count, Reading Time, and Category, followed by Tags and Author.

Edit Document — Update document details and content — allows Title, Summary, Content, Tags, and Author. Category is read-only: it is derived rather than set by hand.

Empty states: No Document Selected — Select a knowledge base and a document to view its details; and No documents for this knowledge base.

Editing a document flips its knowledge base to Needs Update. Re-embed when you are done with a batch of edits rather than after each one.

Sources#

Enlarge

A source is a recurring feed, as opposed to the one-off crawls on the Web Crawler tab. Pick a knowledge base first — Choose a knowledge base to view and manage its sources.

Adding a source

Add New Source:

Sitemap is usually a better choice than Website for a site of any size: it enumerates pages directly instead of relying on link discovery, so nothing orphaned gets missed.

Source cards

Each card shows the URL, a sync status pill (idle, syncing, completed, error) with a live percentage while running, and:

A source is added with a Crawl Depth of 3 and there is no field for it on this page, so a site that buries content deeper than three clicks needs a one-off crawl from the Web Crawler tab, where Max Depth goes to 10.

Actions are Sync and Delete. While anything is syncing the page refreshes every 10 seconds.

Empty: No sources found — Add your first source to start building your knowledge base, with Add Your First Source.

Recent Crawl Jobs

Below the cards, the five most recent crawl jobs for this knowledge base, each with its URL, status, {n} pages processed, a percentage while running, and Duration. Empty: No crawl jobs yet for this knowledge base — Start crawling to see job history here.

For full job history, filters, and per-job details, use the Web Crawler tab.

Web Crawler#

Enlarge

Two tabs: Web Crawler and Statistics.

A banner at the top reports the live connection — Real-time Updates Active with a Live pill, or Real-time Updates Disconnected with Offline — and notes Subscription-driven updates (no polling). If it goes offline you get Real-time updates temporarily unavailable; jobs keep running, but you will need Refresh to see progress.

Starting a crawl

Both are required — Please enter a URL and select a knowledge base. An unparseable address gives Please enter a valid URL (e.g., example.com or https://example.com).

Show Settings reveals:

Delay (ms) is politeness — the pause between requests. Lowering it on your own site is usually fine; lowering it against a site you do not control is not.

Follow external links is off deliberately. Turning it on means the crawler can wander onto sites you do not own and embed their content into your knowledge base, which is both a quality and a licensing problem.

Chunk Size here is not the same field as Chunk Size in Settings. This one applies to the crawl and accepts 100–5000; the one on the Settings tab governs the knowledge base and accepts 0–800. They are separate values with different limits — a source of real confusion.

What the crawler will not read

The crawler follows HTML pages and skips URLs ending in .pdf, .doc, .docx, .xls, .xlsx, .ppt, .pptx, archives, executables, images, and stylesheets.

So a policy document or price list published as a PDF on your website will not be picked up by a crawl, no matter how prominently it is linked. Upload those on the Documents tab. This is the most common reason a crawl "misses" important content.

Jobs

Filter by all, running, pending, completed, failed, cancelled, each with a count, plus Refresh.

A running job shows a progress bar, {n}% Complete, and {processed} / {found} pages, with Stop. Finished jobs show final progress and can be deleted — Delete appears only once a job is no longer running, so stop a runaway crawl first.

Show Details exposes Job ID, Created By, Knowledge Base, Progress, Pages Found, Pages Processed, and the raw Configuration. Quote the Job ID when reporting a problem.

Completion and failure are announced: Crawl job completed! Processed {n} pages. and Crawl job failed for {url}.

Empty: No crawl jobs found.

Statistics

Tiles for Total Jobs, Running, Completed, Failed, Pending, Cancelled, and a Status Distribution breakdown with counts and percentages. Useful for spotting a source that fails consistently.

Settings#

Enlarge

Per knowledge base — pick one first (Select a knowledge base to edit settings.).

Both numeric fields default to their maximum, which is the right starting point — reduce them only to fix a specific retrieval problem, and re-embed afterwards so the change takes effect.

Validation is explicit: Chunk size must be between 0 and 800. and Chunk overlap must be between 0 and 100.

Auto-update sources and embeddings is the setting worth leaving on. With it off, every document change leaves the knowledge base in Needs Update until someone notices.

Language is a hint for processing, not a restriction on what your widget can answer in — Chat supports 50+ languages regardless.

Save Settings commits.

Troubleshooting#

Related#

• Knowledge quick start
• Chat — attaching a knowledge base to a widget
• AI Calls — attaching one to a voice agent
• Integrations — connecting a Shopify store
• Start here — where knowledge fits in first-run setup
• FloGPT for Shopify — catalogue embedding