Adding knowledge
The knowledge base is the content your chatbot draws on to answer questions. Everything you add here is split into chunks, embedded, and made searchable, so the bot can ground its answers in your own material instead of guessing. Open it from Dashboard → Knowledge.
🖼️ [Image] — The Knowledge Base page: stat cards across the top, the filter bar, the Recent Crawl Jobs panel, and the items table.
The three ways content gets in
Click Add Item to open the Add Knowledge Base Item modal. It has three top-level tabs:
| Tab | What it does | Best for |
|---|---|---|
| Text | Paste content directly. | Policies, specific answers, anything that only exists in someone's head. |
| URL | Pull content from the web — Single URL, Crawl from URL, or Import from Sitemap. | Bulk coverage of content you already publish. |
| File | Upload one .pdf, .docx, or .txt. | Manuals, spec sheets, terms documents. |
The bulk web modes have their own deep-dive — see Web crawling & sitemaps. If you want the content to be publicly readable as well as answerable, write it as a documentation page instead: published docs pages sync into this same knowledge base automatically.
Add a text entry
- On the Text tab, enter a Title (required).
- Optionally pick a Category: General, FAQ, Products, Shipping, Returns, Billing, Technical, or Account. Categories don't change how retrieval works — they're there so you can filter a large knowledge base later.
- Paste your Content (required).
- Click Add & Index. The item is chunked and embedded immediately, then enriched in the background.
🖼️ [Image] — The Add Knowledge Base Item modal on the Text tab.
The FAQ category is just a label you apply by hand. It is unrelated to the FAQ module, which discovers question clusters from real conversations rather than from anything you write here.
Upload a file
On the File tab, choose a single .pdf, .docx, or .txt file — maximum 10 MB, one file at a time. Optionally set a Title and Category, then click Upload & Index. Text is extracted from the document, then chunked and embedded like any other item.
A few practical notes:
- Only text is extracted. A PDF that's really a stack of scanned images has no text layer to pull, and will index as empty or near-empty. Check the chunk count on the row afterwards — a
0there means nothing was extracted. - Layout is discarded. Multi-column PDFs, tables, and headers/footers flatten into running text. If a table carries the actual answer, consider restating it as a text entry.
- For long structured documents, the Documentation module has a dedicated importer that turns a
.docxor.pdfinto a tree of separate pages rather than one large blob.
Reading the items table
Each row shows the item's source, category, status, chunk count, and enrichment state. The table paginates 20 items per page.
The chunk count is the most useful number on the row: it tells you the content genuinely made it in. The enrichment badge (Enriched, Partial, Pending, Failed) tells you whether the AI metadata pass finished — see Enrichment & re-indexing.
Filtering
The six stat cards at the top — All Items, Manual, Web Scrape, Files, Products, Docs — double as filters: click one to show only that source. The filter bar underneath adds free-text search, category, enrichment status, and a created-date range.
Row actions
The ⋯ menu on every row gives you:
- Edit — change the title, content, or category. Editing content does not re-embed on its own; follow with Reindex.
- Reindex — re-chunk and re-embed from the current content, and queue a fresh enrichment pass. This is the full rebuild.
- Re-enrich — regenerate the AI metadata only, leaving the embeddings alone. Not offered for product items, which the catalog pipeline enriches instead.
- Delete — remove the item and its chunks.
Select rows with the checkboxes for a bulk delete. Deletion is immediate and the embeddings go with it, so there's nothing to undo — re-add the content if you remove it by mistake.
Limits and what happens when you hit them
Your subscription caps how much you can store, not how many questions the bot answers:
| Trial | With a subscription | |
|---|---|---|
| Knowledge base documents | 100 | 1,000 |
| Total knowledge content | 25 MB | 250 MB |
Limits are enforced server-side, so the dashboard and the API behave identically. An upload or entry that would cross the cap is rejected with a message naming the limit rather than being silently truncated, and a crawl that would cross it stops early and keeps everything it already ingested. Current usage is on the Storage & Infrastructure card of the Billing page. Full table in Pricing & limits.
Only one crawl or sitemap import runs at a time per workspace. Starting a second while one is active is refused with "A crawl job is already running. Please wait for it to complete."
Fixing an answer gap fast
When the bot fails a question, the fastest fix is almost never a bigger crawl — it's one manual Text entry that answers exactly that question, in the words the customer used.
You don't have to guess which ones. Dashboard → Analytics → Content Gaps shows your open unanswered questions and your fallback rate — the questions visitors asked that retrieved nothing. That list is your content backlog, already prioritised by real demand. See Analytics deep-dive.
If you also have the FAQ module, Dashboard → FAQ → Gaps goes a step further: it clusters those questions into recurring themes, and each cluster carries an Add to Knowledge Base action that opens the Add modal for you, plus Archive for the ones not worth answering.
Two habits that keep quality high:
- Write the question into the content. A heading phrased as "How long does delivery take?" gives retrieval a much better target than one titled "Logistics".
- Keep identifiers verbatim. Meaning-based search handles paraphrase well, but exact SKUs, part numbers, and product names still need to appear literally. See How retrieval works.