Four source types
Manual text, PDF, web page and sitemap are managed in one place; each source shows its indexing state.
- Knowledge base
- Source
—PRODUCT
Collects site, PDF, FAQ and table content in four source types; the agent answers only from approved knowledge.
Four document types, automatic refresh and critical-fact approval. The knowledge wizard scores a source from 0 to 100 and prepares the assistant instructions.
—SUMMARY
Collects site, PDF, FAQ and table content in four source types; critical facts go through approval.
01HOW IT WORKS
Each step is one screen in the panel; nothing goes live without confirmation.
Add manual text, a PDF, a web page or a sitemap. Images other than PDF are not accepted.
A source is Ready, Queued, Processing or Failed; when the document count is unknown, 0 is not shown.
Web and sitemap sources refresh every 24 hours; price, return, delivery, warranty and privacy facts go through approval.
02WHAT IT DOES
The three capabilities below are today’s scope; the sample dialogue is fictional and contains no real customer data.
Manual text, PDF, web page and sitemap are managed in one place; each source shows its indexing state.
Web and sitemap sources refresh by default with a 24-hour threshold; unchanged content is not re-indexed.
The “Here is what I found” draft is rule-based; critical items such as price, returns and delivery are not used before approval.
03THE DIFFERENCE IS IN THE DETAILS
21 items, each tied to a rule, setting or automatic behaviour in today’s product. No unsupported number, ratio or promise appears in this list.
A source is Ready, Queued, Processing or Failed; when the document count is unknown it says “Document count unknown”, not 0.
There is no screen to delete or update a knowledge source; if a wrong source is opened, the advice is not to add documents to it.
There are four document types: manual text, PDF, web page and sitemap.
Only PDF is uploaded; images such as PNG/JPG are not accepted, and the file limit is 10 MB.
A PDF with no extractable text is rejected with 400; if the text is very short it is accepted with a “May be scanned” warning, and the agent cannot use that content.
Web and sitemap sources refresh automatically by default with a 24-hour threshold; unchanged content is not re-indexed.
A single document is at most 180,000 characters; the 8,000-per-page sitemap limit applies only when the broad-coverage setting is on.
“Here is what I found” is rule-based and does not use a model; it extracts address, phone, email, hours, price, social and Q&A groups.
Price, return, delivery, warranty and privacy facts are not used before approval; checking the box is not enough — “Add approved items to the knowledge base” is required.
A manually edited item is marked “manually edited” and gets 100/100 confidence; an unanswered question cannot be approved, and a new scan with a different value offers “Keep my value / Use the new value”.
If there is no coverage test, no fake percentage is shown; the panel says “not tested yet”.
The knowledge wizard has four steps: Get to know you, Source type, Details, Summary and test.
The wizard has 16 ready sectors and “Other”; Other uses the nearest sector or a general structure.
Goals are multi-select: present products, recommend, collect contact details, appointments, orders/shipping and FAQ.
The source score runs from 0 to 100: 75+ is sufficient, 50–74 can improve, below 50 needs work.
At 75 or above the wizard says “Sufficient, continue”; adding another source is optional.
The wizard does not claim “12 pages read”; it shows the real PENDING/PROCESSING/READY/FAILED state.
If a scanned PDF is accepted, it is rolled back and the panel says “not added”.
The test uses the real agents/:id/test address and does not keep a conversation record.
A manually edited instruction is not overwritten; the panel says the current instructions were preserved, and a new suggestion appears only under Advanced.
If the agent is not connected to the target source, the panel warns “The agent is not connected to this knowledge source”.
04CONTROL AND SECURITY
These are this surface’s current security behaviours; no general “GDPR compliant” certificate is claimed.
Price, return, delivery, warranty and privacy facts do not enter answers before approval; approval takes two steps.
A wrong source cannot be deleted; verify content before opening a source and do not add documents to a wrong one.
A PDF with no text is rejected; a low-text PDF is accepted with a warning, and the agent cannot use that content.
If it has not been tested, the panel does not invent a percentage; it says “not tested yet”.
05WHICH PLAN
Document and source counts depend on the plan: 3 documents in trial, 5 sources/5 documents on Starter, 10/50 on Growth and 50/500 on Business.
All amounts are monthly and exclude VAT. See the plans page for the current list and calculator. Plans and calculator →
06FAQ
+RELATED POSTS
Posts from the same topic cluster that complete this page. Knowledge base posts.
A practical comparison for SMEs: when a call center makes sense, when an AI voice assistant does, and how volume, scope, handover, cost and data should decide.
AI support plans count messages in some places and conversations in others. Five questions to ask and how the two models compare.
Five common multi-channel customer communication mistakes — tone, handover, stale knowledge, ownership and measurement — and how to fix them.
Try free for 14 days without a credit card; we will do the first setup together.