Manage knowledge sources: preview, edit and retrain
Manage an intoCHAT agent's knowledge: training status, preview, edit, retrain, delete, priority, usage, conflicting facts, grouping by site and folder, ZIP export and auto-resync.
Everything your agent knows is listed under Knowledge sources on the Knowledge tab, one row per page, file, text snippet, Q&A pair, Notion page, Google Drive file or Intercom article. Imported helpdesk tickets are the exception: they are shown as one row with a count on the Helpdesk tickets card below the list (see Helpdesk tickets and Intercom). This page explains what you can do with each source and when changes take effect.
Training status
Each source shows a status:
| Status | Meaning |
|---|---|
| Pending | Saved, but not trained yet. The agent doesn't use it. |
| Training | Being split and indexed right now. |
| Trained | Used for answers. |
| Failed | Couldn't be trained. Hover over the badge to see why. |
The bar at the top of the tab sums this up, for example "3 sources waiting to be trained", and shows how many of your plan's characters the agent uses. Click Train agent to train everything that is pending, or Retry failed when only failed sources are left.
Training runs on intoCHAT's servers. After you click Train agent, you can switch tabs or close the page, and training continues. When you come back, the Knowledge tab shows its progress. If a run is interrupted, it picks up again when you open the Knowledge tab, and a daily check also restarts interrupted runs. Finished sources stay trained. If training stops after repeated errors, click Train agent to try again.
Only adding knowledge needs the page open: wait until new files and pages show as added before you leave.
Find sources
Use Search sources… to search by name, address or a Q&A pair's other phrasings, the type menu (All types, Website, Files, Text, Q&A, and Notion, Google Drive or Intercom once you have such sources) and the status buttons, such as Pending (3) or Failed (1).
Group sources by site and folder
In the Arrange sources menu next to the type menu, choose Group by site and folder. List switches back.
- Website pages are grouped by site, then by the site's top-level folder, such as
/blogor/docs. Pages at the top of the site are under/. A site with only one folder lists its pages straight under the site. - Every other source is grouped by type: Files, Text, Q&A, Notion, Google Drive and Intercom articles.
- Each heading shows how many sources the group holds and their characters, for example "24 sources · 81.2K chars". Click the arrow to collapse or expand it.
- Search, the filters and Least used still apply inside each group.
Each group heading has these controls:
- Checkbox: selects every source in the group. It shows a partly selected state when only some are selected. Then use Export, Re-index or Delete for the selection.
- Cloud (refresh from the website): site and folder groups only. After you confirm with Refresh pages, every page in the group is read again and the pages whose text changed are retrained. It uses a crawl from your plan, and one crawl reads up to 600 pages. A group with more pages can't be refreshed at once, so refresh its folders one by one. A page read within the last hour may come from a cached copy. The refresh stops at your crawl or character limit, and a message sums up how many pages changed, read the same or couldn't be read.
- Bin: deletes every source in the group once you confirm with Delete … sources. This can't be undone.
How often a source is used
Each row shows how many of the agent's replies in the last 30 days used the source:
- "Used in 12 replies (30 days)": that many replies had at least one excerpt from this source among the knowledge the agent was given.
- "Not used in 30 days": no reply in that time had an excerpt from it.
Q&A pairs are counted like any other source. Saved Playground chats count too; temporary chats aren't stored, so they don't. The figures start from when intoCHAT began recording which knowledge each reply used. Until the agent has such a reply, the list shows "How often each source is used appears here as new replies are recorded." instead.
Once there are figures, a sort menu next to the type menu offers Default order and Least used. Least used lists sources that weren't used first, then the ones used in the fewest replies, then the ones used longest ago.
"Used" means the source's text was given to the agent for that reply, not that the reply quoted it. A source that is never used may cover topics visitors don't ask about, or be hard for search to find. Preview it before you delete it, in case it holds facts you still need. To see which questions your knowledge doesn't answer yet, use Questions to improve; see Text and Q&A.
Conflicting facts
On Pro and Enterprise, the Conflicts card on the Knowledge tab lists facts that two of your sources disagree on, such as an old price on one page and a new one in a Q&A pair. Open source A or Open source B opens the source in this list: text snippets and Q&A pairs for editing, other sources in their preview. Once the corrected source is trained, the finding is resolved by itself. See Conflicts.
What you can do with a source
Each row has these icons:
- Star (Mark as priority): the source ranks above other matching knowledge and wins when sources disagree. It applies to the next answer, without training. Q&A pairs don't have a star because they already rank first. Marked sources show Priority.
- Eye (Preview the text): shows exactly the text the agent learns from the source, Formatted or as Plain text. You can search it, Copy it, or Download it as a Markdown (
.md) file. For a web page, Open page opens the original. - Pencil (Edit): text snippets and Q&A pairs only. Change the title, content, question, other phrasings or answer, then Save. The source goes back to Pending; train to apply the change. A question that another pair already has is refused.
- Cloud (Refresh from website): website pages only. Reads the page again and, if its text changed, saves and retrains it right away. It uses a website crawl, like adding a page.
- Arrows (Re-index saved text): trains the source again from the text intoCHAT already has, without fetching anything.
- Bin (Delete this source): removes the source and its passages.
Notion pages and Google Drive files have no edit or refresh icon. To update one, import it again or let auto-resync check it; see Notion and Google Drive, where Sync now checks all of an agent's Drive files at once.
Tick several rows to Export, Re-index or Delete them together. To add, update or download many Q&A pairs at once, use the import and export under Q&A; see Text and Q&A. Delete all at the top removes every source once you confirm with Delete everything. This can't be undone.
When changes take effect
| Change | Takes effect |
|---|---|
| Mark or unmark as priority | With the next answer |
| Delete a source | Immediately |
| Edit a text snippet or Q&A pair | After training. Until then the agent keeps using the previous version. |
| Refresh a website page that changed | When its automatic retraining finishes |
| Auto-resync finds a changed page | When its automatic retraining finishes |
| Upload a new version of a file | After training. Until then the agent keeps using the previous version. |
| Add a source | After training |
Deleting sources also frees characters for new ones.
Updating website pages and files
Re-index saved text doesn't fetch anything new, and website pages and files can't be edited in place. Website pages, Notion pages and Google Drive files can be checked for you on a schedule with auto-resync, and helpdesk tickets and Intercom articles with their own Sync automatically setting (see Helpdesk tickets and Intercom). Otherwise, update the sources yourself when your prices or policies change:
- A web page changed: click Refresh from website on its source. A page read within the last hour may come from a cached copy. To update many pages at once, use the cloud icon on a group in Group by site and folder, or add the pages again after Find pages and train. See Website crawl.
- A Notion page changed: import it again. See Notion.
- A Google Drive file changed: click Sync now on the Drive option, or pick the file again. See Google Drive.
- A file changed: upload the new version under the same file name, then train. It replaces the old text; an identical file is skipped. See Files.
- A snippet or Q&A pair changed: edit it with the pencil icon and train.
Auto-resync
The Auto-resync card on the Knowledge tab reads your website pages, Notion pages and Google Drive files again on a schedule and retrains only the pages whose text changed. It is off by default.
- Turn on Check pages automatically.
- Under How often, choose Daily or Weekly. Each change is saved at once.
What a run does:
- It reads website, Notion and Google Drive sources again. Uploaded files, text snippets and Q&A pairs are never touched.
- A Google Drive file is read again only when it was modified in Drive since the last check. A file that was deleted, moved to the trash or that intoCHAT may no longer open shows the reason, and its saved text stays.
- It compares each page with its saved text. Text that repeats across your site, such as menus and footers, is ignored in the comparison, so it doesn't count as a change.
- It saves the new text of changed pages and retrains them by itself. Unchanged pages are only marked as checked.
- It never deletes a source. If a page can't be read, for example because it was removed from your site, the saved text stays and the source shows the reason.
- If a page's new text would take the agent over its character limit, the page keeps its saved text and shows "The new text would go over your plan's character limit, so the saved text was kept."
- Pages checked longest ago go first. A website page read within the last hour may come from a cached copy.
- Website pages are read with the agent's crawl settings: its scraping country, and in slow mode 5 pages at a time, 2 at once, with a pause between. In slow mode a run can read fewer than 30 pages in its time; the rest are checked on the next run.
Website crawls
- Each run that reads website pages uses 1 website crawl from your plan and covers up to 30 pages.
- A site with more pages is covered over several runs, one a day, each using a crawl. With 90 pages, one round takes three runs and three crawls.
- For up to 30 pages, daily checks use about 30 crawls a month and weekly checks about 4 to 5. Free, Starter and Pro include 1, 9 and 20 crawls a month, shared with adding pages, so daily checks of a website use up these plans' crawls.
- Without crawls left, website pages aren't read and wait until your crawls start over. The card says how many pages weren't checked, and those sources show "Not checked: Out of website crawls for this period (saved text kept)".
- Notion pages and Google Drive files use no crawls. A run reads up to 30 Notion pages and up to 30 Drive files.
When runs happen
- Runs happen at most once a day. The daily job starts around 05:00 UTC; the card shows that time in your time zone. When many agents are due, the job continues in further steps, so on busy days a run can start later.
- With Daily, a page is due when its last check is about 20 hours old. With Weekly, about six and a half days.
- The card shows Last run, Next run (an estimate) and Last run found, for example "12 checked · 2 changed · 0 failed".
Check now
Check now runs a check right away and waits for it, which can take up to a minute. It checks all pages, not only the ones that are due, within the same limits: up to 30 website pages, using 1 crawl if there are website pages, and up to 30 Notion pages and 30 Google Drive files, those checked longest ago first. Pages it doesn't reach are checked on the next run. It also works while auto-resync is off.
You can use Check now once every 10 minutes per agent; until then the card shows "Available again at …". After a check, the next scheduled run waits at least 20 hours.
Status on each source
Website, Notion and Google Drive sources that have been checked show a line "Checked … · changed …": the date of the last check and the date the text last changed. Hover over a date to see the exact time. When the last check failed, the line reads "Check on … failed:", followed by the reason and "(saved text kept)". When the reason suggests the site turned the crawler away, such as a 403 or a bot check, a second line links to Sites that block crawlers.
Export sources
You can download your sources' text as a .zip file:
- Click Export all above the list to export every source.
- Or tick sources, or a group's checkbox, and click Export in the bar that appears.
The file is named like knowledge-sources-2026-10-07.zip. It holds one file per source, in a folder per type: website/, files/, text/, qa/, notion/ and google-drive/.
- Each file holds the saved text the agent learns from, as
.txt. For an uploaded file, that is the extracted text, not the original document. - Q&A pairs are
.mdfiles in the formQ: …andA: …, with each other phrasing on its ownQ:line. manifest.csvlists every source with the columns Name, Type, URL, Characters, Status and File.
One export holds up to 50 million characters of text, about 50 MB. For more, select fewer sources and export them in parts.
When to use Re-index
Use Re-index saved text on a source that failed, for example after a temporary error, or select several sources and click Re-index to train them again. For content changes, refresh, re-upload or edit the source as described above.
Good habits
- Preview new sources before training. A page that only captured its menu, or a PDF table that came out garbled, is easy to spot there.
- Remove outdated or duplicate sources instead of hoping the agent picks the right one.
- If the agent keeps quoting the wrong source, mark the correct one as priority or add a Q&A pair.