Knowledge base

What the knowledge base is forCopy Anchor Link

When the AI answers a customer, it may only use what you have given it: the knowledge base, your policy numbers, skills, synced product data, and a signed-in customer's own orders (the full picture is in How the AI answers). The knowledge base is the biggest of these. It lives under Knowledge base (知识库) in the console sidebar.

Everyone on the team can view it. Only owners and admins can add, edit, delete or restore older versions.

Add (添加) offers five ways in:

Every entry has a status: Indexing (索引中), then Ready (可用), or Failed (失败). Ready means a customer's question can already find it; a new entry usually takes a little while. The list refreshes by itself.

ArticlesCopy Anchor Link

• Title up to 200 characters, body up to 200,000 characters.
• Public link (公开链接) is optional. If you fill it in, the AI adds the link after its reply whenever it cites the article.
• Write in complete sentences a customer can follow. The AI quotes what you wrote and is instructed not to say anything your sources do not.

Q&ACopy Anchor Link

A Q&A entry is a question, other ways of asking it, and an answer. Customers who phrase the question differently still reach it, so list the common variants.

• Question 2–500 characters; up to 20 other ways of asking (duplicates, and any identical to the question, are removed); answer up to 20,000 characters.
• Use Edit (编辑) on an entry to change it. Editing keeps the entry's ID, so answers that already cited it still link to it.

Importing from a spreadsheetCopy Anchor Link

Excel (.xlsx) and CSV are supported. Put headers in the first row; Chinese and English headers are both recognised:

• Without recognisable headers, two columns are read as question and answer, three as question, other ways to ask, and answer.
• Put several variants in one cell separated by line breaks, |, ； or ;.
• You see a preview first. Rows with problems (missing question, missing answer, too long) are listed by row number and skipped.

Duplicates: an entry whose question matches an existing one (case-insensitive) updates it; an identical entry is left alone; everything else is created. When the import finishes you see how many were created, updated and unchanged. Importing the same file again after an interruption does not create duplicates.

Uploading filesCopy Anchor Link

• Up to 25 MB per original file; up to about 3.5 MB of text after conversion.
• Files are converted to text and indexed. The original is kept.
• The formats you can upload are the extensions listed after "Supported" (支持) in the upload dialog. That list comes from the conversion service itself and we refresh it daily, so formats it adds work without waiting for us. Formats we have verified include PDF, Word (.docx), PowerPoint (.pptx), Excel (.xlsx), CSV, HTML pages, images and ODT.
• In spreadsheet files (xlsx, csv and similar), each row becomes one line of "column: value · column: value", so even when only one row is retrieved, the AI knows what each value means.
• The title is the document's first heading, or the file name if it has none.
• Scans (PDFs that are only images) contain no readable text; you will be told no text could be read. Images embedded in PDF and Word files are not read.

There are two ways to update a file. Both keep the entry's ID, so the AI's earlier citations keep working:

1. Upload a file with the same name again; the old content is replaced.
2. Use Replace file (替换文件) on the entry in the list and pick the new version. The name may differ, for example price-list-v2.pdf replacing price-list.pdf.

If a new file has exactly the same content as an existing file under a different name, the upload is refused and you are pointed to Replace file on that entry.

Importing a whole help centreCopy Anchor Link

Choose Add → Import website, enter the address of your help centre or FAQ page (it must start with https://), set Maximum pages to read (最多读取页数, default 300), and click Start reading (开始读取).

Which pages are read

• Pages on the same site as the start address (with and without www count as one site), under the start address's path.
• Pages are found through the sitemap (sitemap indexes and .gz are supported) and the links on each page.
• robots.txt is respected: disallowed pages are not read, and if the start address itself is disallowed the import fails. A crawl-delay is honoured, up to 10 seconds per page.
• Pages whose canonical points elsewhere, or that redirect to a page already read, count as duplicates.
• FAQ structured data on a page (FAQPage) becomes Q&A entries.
• Pages with no readable text over a plain request (JavaScript-only shells, or pages that refused the request) are read again with a browser. Either way the request identifies itself as PurserBot; it never pretends to be an ordinary visitor.
• If the site blocks us with a bot check or a 403, you are told to allow PurserBot in the site's firewall and import again.
• If the start address redirects somewhere else, the error says where, so you can import from that address instead.

The first import asks you to confirm the scope

The first time a site is read, the import stops and shows that reading is done and it is waiting for you to confirm what to import. Click Confirm import scope (确认导入范围):

• Choose languages to import (pages without a language tag are always imported).
• Choose sections; you can expand each section to see its pages. Leave out blogs, job listings, legal terms and anything else the AI should not quote.
• "Pages not imported" lists the reason for each page: no body text, blocked by the site, disallowed by robots.txt, failed to open, page does not exist, duplicate.

Click Import N pages (导入 N 页) to go ahead or Don't import (不导入) to cancel. If you do not confirm within 7 days, the import is cancelled. Your choices are remembered and later automatic updates follow them.

Pages read per import

Automatic updates

• Each site under Imported websites (导入的网站) has an Auto-update (自动更新) switch, on by default.
• About every 6 hours the sitemap is checked. If it changed, or 7 days have passed since the last read, the site is read again. You can also click Update now (现在更新) at any time.
• Only pages whose content changed are imported again.
• A page leaves the knowledge base only after two reads in a row could not find it. A read that stopped at the page limit before covering the whole site does not count.
• After every help-centre import, a content conflict check runs automatically (see below).

Draft and publishedCopy Anchor Link

Every change you make to the knowledge base goes into the draft first.

• Test chat (试聊) always uses the draft, so you can try a change straight away.
• Before your first publish, customers get the draft too, and changes are live immediately.
• After your first publish, customers only get the latest published version. Draft changes reach customers when you publish again.
• A publish only includes entries whose status is Ready; entries still indexing are left out of that publish.

Publishing, rollback and the checks before a publish are covered in Test chat and releases.

Version historyCopy Anchor Link

Open an entry from the list; its full text and Version history (历史版本) appear on the right.

• Before each change (an edit, a re-upload, a site update) the previous text is saved as a version. Nothing is saved if the text did not change. The latest 10 versions of each entry are kept.
• Compare (对比): red is what the old version had and the current one does not; green is what is new now.
• Restore this version (恢复这个版本): puts the old text back. The current text is saved as a version first, so a restore can itself be undone.
• Deleting an entry deletes its versions too.

Retrieval testCopy Anchor Link

The Retrieval test (检索测试) card shows what the AI would be given for a question:

1. Type something a customer might ask and click Search (检索).
2. Choose what to search: "What customers see now" (顾客现在看到的) — the published version, or the draft if you have never published — or "Draft (used by test chat)" (草稿（试聊用）).
3. Up to 8 passages are listed with the entry they come from and a relevance score; open one to see the entry.

The retrieval test only searches. It calls no model and costs nothing. If nothing is found, the AI would have nothing to go on for that question and would hand the customer to a person; add the answer as a Q&A entry.

Citation statsCopy Anchor Link

Each entry in the list shows figures for the last 90 days:

• Cited N (引用 N): how many AI replies cited it (only citations the reply actually used). Hover to see when it was last cited.
• Corrected N (被改 N): how many of those replies a person changed substantially before sending (more than 10% of the text), or marked as wrong.

Filters above the list:

Other tools on the same pageCopy Anchor Link

• Policy numbers (政策数字): return window (days), who pays return shipping, restocking fee (%), refund method, exchanges allowed, processing time (business days), free-shipping threshold, ships internationally, warranty (days), support hours. When the AI states days, fees or thresholds, it takes them only from here, never guessed from an article. The refund rules use the return window and warranty from here too. Leave a field empty if it does not apply.
• Knowledge gaps (知识缺口): Find gaps (找缺口) looks at questions from the last 30 days that customers kept asking and the AI could not answer (including imported ticket history), and drafts an article for each group.
• Content conflicts (内容冲突): finds articles that contradict each other or your policy numbers (an old page saying 14 days, a new policy saying 30). It runs after every help-centre import, or when you click Check now (立即检查).

Plan limitsCopy Anchor Link

• Imported web pages do not count toward knowledge entries; they are limited only by pages per import.
• The published copies a release makes are not counted again.
• Adding beyond the limit shows the plan's maximum and which plan lets you continue. Downgrading never deletes what you already have.

See Plan and billing for plan details.