Knowledge base: give the assistant your rules, and show the source

In user language: for people who want the assistant to answer from the team’s own materials. Write policies, processes, and project notes as how-to articles; keep historical Word / PDF as original files. When you ask, the assistant pulls the relevant passages, lists sources under the reply, and you can open the original. In the product this is the knowledge base (materials for answering in conversation), alongside the standards library and the asset library.

For the user-facing notes, see [help/knowledge-layers/three-layers.md](/docs/help-knowledge-layers); for the mechanism source of truth, see [docs/core-mechanisms/智能体调用知识文档的方式.md](/docs/mech-knowledge).

Date: 2026-08-30

The takeaway (read this first)

Cadau does not stuff an entire policy book into the assistant at once. You maintain a knowledge base; the assistant takes relevant passages for this one question, then answers from them. Ordinary Q&A uses current material by default; only when you clearly ask about “the previous version / an old notice / the archive” does it look at archived files.

The same entry has two parts:

• How-to articles: keep writing Markdown — current metric, FAQ, processes.
• Source archive: upload historical Word / Excel / PPT / PDF / images; preview the original; mark current or archived. The assistant retrieves extracted text. Originals are not rewritten as wiki, and they do not go into the asset library.

In one sentence: how-to articles hold the current metric; originals keep the true source; conversation only carries relevant passages, and you can see where they came from.

flowchart TD
  Q["You say one sentence to the assistant"] --> L{"Which layer does this sentence belong to"}
  L -->|"how to use the product"| G["System knowledge base\nhelp assistant"]
  L -->|"team policy / project wiki"| W["Workspace knowledge base\nshared by every assistant in this workspace"]
  L -->|"this assistant’s dedicated metric"| U["Agent knowledge base\nthis one only"]
  G --> R["Pull relevant passages"]
  W --> R
  U --> R
  R --> A["Answer"]
  A --> C["Sources listed underneath\nopen to preview the original"]

Who this is for, and what it solves

It fits teams that already have policies, notices, and project notes, and want the assistant to speak from your own materials rather than invent them.

Typical situations:

• A new colleague asks “how do I take annual leave” or “which path does expense follow”, and you want the answer to match the current policy, with the original openable to check;
• HR, legal, and the project team each have an assistant, but the same workspace’s policy is maintained once — you do not copy it into every assistant;
• You still have years of Word / PDF notices; you do not want to rewrite them as wiki first, but you want the assistant to answer from extracted text;
• You worry the assistant will pour an entire policy into context and fill the window during small talk; you only want the passages that match this question.

For people getting work done: pick the right assistant and ask; if there are sources under the reply, you can open them. For people who maintain materials: team-shared writing goes in the workspace layer; one assistant’s wording goes in that assistant’s own layer; after edits, generate the index so the next round can find it.

Three layers: where to write, who sees it

In the same workspace, different members’ assistants all read the workspace layer. Personal material and how a given assistant is addressed externally do not belong in the workspace layer.

Entries:

1. System layer: Admin console → System knowledge base.
2. Workspace layer: select a workspace first → Workspace collaboration → Manage workspace knowledge base. You can also tell the help assistant “open the workspace knowledge base”.
3. Agent layer: Agents → My agents → select the instance → Knowledge documents.

With no workspace selected, you can only use system-layer help notes (creating or renaming a workspace can still be done for you). You must enter a workspace to maintain that workspace’s knowledge.

How-to articles and source archive

The workspace and agent knowledge pages are the same entry, two kinds of content, not a separate plugin.

How-to articles

Current metric written for the assistant to read. Fits FAQ, process summaries, project wiki. You can paste images and short explainer clips of a few to a dozen seconds. After you edit the body, the next round can find the new content; please Generate index with AI as well, so directory summaries are not stale.

Bulk migrate: at the current browse location you can upload a folder / archive or download an archive (Markdown and the index).

Source archive

Fits historical policy, old notices, scans. After upload you can preview the original; each file is marked current or archived. The assistant retrieves extracted text. It does not rewrite the original as a Markdown article in the knowledge tree, and it does not go into the asset library (the asset library holds reusable sheets and templates).

If parsing fails, retrieval test clearly says there are no passages yet; it does not pretend it understood the file.

What actually happens in conversation

You send one sentence; the assistant takes relevant passages (not the whole directory). Under the reply it lists sources: how-to articles by title, source-archive files by file name; open to preview.

When the workspace metric conflicts with one assistant’s dedicated notes: if this assistant states an overlay, that wins; otherwise the workspace wins; if neither covers it, it should say it does not know, not invent.

After you edit a how-to article or ingest an original, the next round can find it. No restart, and you do not paste the full text into the conversation again.

Try retrieval first: retrieval test

The knowledge base offers retrieval test: type a sentence the assistant would hear, and immediately see the extracted notes. How-to articles pick a topic from “Generate index with AI” then open the body; source-archive files match on extracted text.

It only tests materials in the directory; it does not search past conversations. If there is no match, it says no related content was retrieved. If this sentence is looking up something said in a past conversation, the test also explains that the directory will not extract it — ask the assistant in Messages to look through past conversations.

Use it to check:

• Whether this sentence hits the policy you just wrote;
• Whether the hit is a how-to article or an original file;
• Whether you already generated the index with AI after editing.

After you maintain materials, try one sentence here, then go back to Messages and ask the sentence a real person would ask — cheaper than renaming titles by feel.

How this splits with the things next to it

The knowledge base serves conversation Q&A and the team wiki. Several neighbours have different entries; do not mix them.

• Regulations with versions that need human review → standards library; do not drop them into knowledge documents and pretend they are reviewed clauses.
• Templates for colleagues to download and fill → asset library.
• “From now on always answer in bullets” → say remember, which goes to memory.
• “Zhang San’s phone, August leave” → memos, not the team wiki.

The help assistant mainly reads the system layer and answers how to use the product; for team policy, select an assistant in the workspace and ask again.

What you can say to the assistant

Say what you want done, and it acts; ask “how / how do I maintain knowledge” and it explains the steps.

With no workspace selected: questions about how to use the product are answered by the help assistant from system knowledge; for team policy, select a workspace first, then say it again to a work assistant.

What to do after you edit

1. Edit the body or ingest an original → the next round of conversation can use the new passages.
2. Generate index with AI (you can force regenerate even if the body did not change) → directory summaries and topic keywords catch up with the body, so retrieval is more stable.
3. Optional: retrieval test — paste a sentence a real person would ask, and see whether the extracted notes are right. After editing a how-to article, generate the index with AI, then try retrieval.

At the system layer, tap “Rebuild index with AI” in the admin console. Workspace and agent layers generate on their own knowledge pages.

Index entries can say “use this article only when the question or attachment path is related”. Do not put that kind of restriction on general overviews or account notes, or everyday Q&A will fail to retrieve them.

Data dictionary: one table, one file

Table structure for the assistant to read should be one how-to article per table (or view), e.g. 数据字典/eaemp.md. Inside the article, table name, purpose, and a field table are enough.

• An article of about 5,000 characters or less (typical field notes for one table fit) is retrieved as one passage, so “what fields does eaemp have” does not only pull half a table.
• Extra-long tables can be grouped under ## (e.g. “Primary key and status”, “Payroll fields”); the system merges sections that are too short, and tries not to cut in the middle of a table.
• Do not put a dozen tables in one file, or asking about one table will drag in others, or the overlong file will be truncated.

When not to force knowledge documents

• Personal preferences you want every future conversation to follow — say “please remember”; do not write them into team knowledge.
• Regulation originals and standards that need human review — go to the standards library.
• Blank sheets for colleagues to fill — go to the asset library.
• Employee files, leave, invoices — go to memos.
• What this conversation is doing right now — write it in the topic; the conclusion after it is done goes in review.
• Only asking how to click around Cadau — use the help assistant (system layer); do not mistake an operations manual for workspace policy.

Purpose goes in the topic, the conclusion after the work is done goes in review, habits across conversations go in memory, records you look up when needed go in memos, team metric goes in the knowledge base.

Related reading

• Three layers of the knowledge base — what to write at system / workspace / agent
• Who can maintain which layer — entries and generating the index after edits
• How agents call knowledge documents — how passages enter conversation
• Standards library — true source of regulations; do not mix with wiki
• Asset library — reusable templates and sheets
• Using agent memos — records you look up when needed
• Product introduction — where knowledge documents sit in the product