@nocobase/plugin-ai 2.2.0-alpha.1 → 2.2.0-alpha.10

package/dist/ai/docs/nocobase/ai-employees/knowledge-base/knowledge-base/settings.md

@@ -0,0 +1,77 @@
+---
+pkg: "@nocobase/plugin-ai-knowledge-base"
+title: "Settings"
+description: "Modify basic information, vector store, and default segmentation settings on the Settings page of a knowledge base, and understand the relation between knowledge-base-level and document-level segmentation settings."
+keywords: "knowledge base settings,segmentation settings,Chunk size,Chunk overlap,vector store,NocoBase"
+---
+
+# Settings
+
+## Open the Settings page
+
+After opening the knowledge base detail page, click Settings in the left sidebar. This page modifies basic information, file storage, vector store, and default segmentation settings of the current knowledge base.
+
+![](https://static-docs.nocobase.com/20260617005832.png)
+
+## Basic information
+
+Local knowledge base settings include:
+
+| Setting | Description |
+| --- | --- |
+| Key | Unique identifier of the knowledge base. It cannot be changed after creation |
+| Name | Knowledge base name |
+| File storage | Where documents and segment files are stored. It cannot be changed after creation |
+| Vector store | Vector store used by the current knowledge base |
+| Description | Knowledge base description |
+| Enabled | Whether the current knowledge base is enabled |
+
+:::tip Vector store changes
+
+After you modify Vector store, NocoBase asks for confirmation when saving. After the vector store changes, existing documents must be vectorized again before vector data can be written to the new vector store. Choose Save and vectorize to save the settings and vectorize immediately, or choose Save only to save the settings and later run Vectorization manually on the Documents page.
+
+:::
+
+![](https://static-docs.nocobase.com/20260617005951.png)
+
+## Default segmentation parameters
+
+The Settings page includes three segmentation parameters:
+
+| Setting | Default | Description |
+| --- | --- | --- |
+| Split document | Enabled | Whether uploaded documents are automatically split into multiple segments |
+| Chunk size | `6000` | Maximum character count of each segment |
+| Chunk overlap | `1200` | Overlapping character count between adjacent segments |
+
+Chunk overlap cannot be greater than or equal to Chunk size. If the value is invalid, the backend normalizes it to keep overlap smaller than a single segment.
+
+## Relation to document-level segment settings
+
+These settings are knowledge-base-level defaults. They mainly affect documents uploaded or regenerated later.
+
+After a document enters the knowledge base, it keeps its own segmentation parameters. On the Documents page, click Segments for a document, then open Segment settings to set Split document, Chunk size, and Chunk overlap for that document and run Resegment.
+
+The relation between the two places is:
+
+- Segmentation parameters in Settings define the default strategy for the knowledge base and are suitable for controlling new documents consistently
+- Segmentation parameters in the Segments dialog only apply to the current document and are suitable for handling a document that is too long, too short, or structurally special
+- Existing generated segments are not automatically rebuilt after knowledge-base-level defaults change
+- To apply new parameters to an existing document, open its Segment settings and run Resegment
+
+:::warning Note
+
+Resegment discards manually edited segment content and related questions in that document. Changing knowledge-base-level defaults does not immediately overwrite existing segments; only document-level Resegment rebuilds segments.
+
+:::
+
+## When to adjust parameters
+
+Default parameters cover most plain text, Markdown, Word, and PDF documents. Adjust them only in these cases:
+
+- Segments are too long and matched results include too much unrelated content: reduce Chunk size
+- Segments are too short and context is incomplete: increase Chunk size
+- Sentence breaks are obvious between adjacent segments: increase Chunk overlap
+- You want the whole document to be one retrieval unit: turn off Split document
+
+After saving settings, newly uploaded documents use the new default parameters.

package/dist/ai/docs/nocobase/ai-employees/knowledge-base/rag.md

@@ -19,7 +19,7 @@ Select the AI employee for which you want to enable RAG, click the `Edit` button
 In the `Knowledge base` tab, turn on the `Enable` switch.
 
 - In `Knowledge Base Prompt`, enter the prompt for referencing the knowledge base. `{knowledgeBaseData}` is a fixed placeholder and should not be modified.
-- In `Knowledge Base`, select the configured knowledge base. See: [Knowledge Base](/ai-employees/knowledge-base/knowledge-base).
+- In `Knowledge Base`, select the configured knowledge base. See: [Knowledge Base](./knowledge-base/).
 - In the `Top K` input box, enter the number of documents to retrieve, the default is 3.
 - In the `Score` input box, enter the document relevance threshold for retrieval.
 

package/dist/ai/docs/nocobase/ai-employees/scenarios/company-background-research.md

@@ -0,0 +1,125 @@
+---
+title: "Workflow + AI for employees to complete company background research automation"
+description: "Through company information forms, background investigation records, workflows, and AI employees, a company background investigation process can be automatically triggered, retained, and supported for manual review."
+keywords: "NocoBase, AI employees, workflow, company background research, due diligence, automation, business practice"
+---
+
+# Workflow + AI for employees to complete company background research automation
+
+In NocoBase, you can turn company background research into a trackable automated task flow. Business staff still work in the familiar company information page, while workflow and AI staff are responsible for completing background information, recording the processing process, and saving each generated report.
+
+![](https://static-docs.nocobase.com/202607121806356.png)
+
+This scenario is suitable for dealing with a common problem: company background information is not a static field that ends after being entered once. Public information will change, regulatory events will occur, and the status of cooperation will be constantly adjusted as business progresses. If you only rely on manual supplementary recording on a regular basis, it will be easy to miss; if you directly let AI cover company information, it will be difficult to explain "how this judgment came about." The approach here is to separate and save the current data and the research process - the company record saves the version being used by the business personnel, and the background check record saves the status, output and history of each AI survey.
+
+## Let’s look at the two tables first
+
+The company information form provides the basic information of the research object, and the background investigation record form is responsible for undertaking each research task. One saves the currently available information, and the other saves the processing process and historical results.
+
+### `companies`: Company information table
+
+| Core fields               | effect                                                           |
+| ---------------------- | -------------------------------------------------------------- |
+| Company name           | The main identifying information of the research object.                                   |
+| Website                | Provide official website clues to reduce misjudgments caused by companies with the same name or abbreviation.                   |
+| Address                | Assist in determining regions, entities and business scope.                                 |
+| Company type           | Mark business relationships such as customers, suppliers, partners, etc. to facilitate subsequent judgment and follow-up priorities. |
+| Background information | Save the company background report you are currently using and use Markdown to render structured content. |
+
+### `background_check_tasks`: Background check record form
+
+| Core fields                  | effect                                                                               |
+| ------------------------- | ---------------------------------------------------------------------------------- |
+| Company ID / Company name | Record which company this survey is for to facilitate task execution and historical review.                                 |
+| Status                    | The flow of marking tasks from `pending` to `processing` and `completed` is also the basis for preventing repeated triggering. |
+| Research report           | Save the complete research report generated by AI this time.                                                   |
+| Summary                   | Save AI's summary of the research process, risk points, and information to be supplemented.                                     |
+| Previous background       | Save the old version before writing back, supporting historical tracking and comparison of old and new reports.                                   |
+
+![](https://static-docs.nocobase.com/202607121807627.png)
+
+## Enter the research process from company information
+
+The company list is the most familiar entrance for business people. You can see the company name, official website, company type, contact person, email and other information on the page. After entering a company, business personnel can view the current background report or initiate a new background investigation.
+
+After entering the editing page, "Background information" is displayed using the Markdown editing component. The AI-generated content is not a short summary, but a structured report that can be read, copied, and continued to be maintained. Business personnel can still modify it manually, but each result generated by AI will leave a corresponding history in the background check record.
+
+![](https://static-docs.nocobase.com/202607121807450.png)
+
+In this way, the page still looks like an ordinary company data maintenance interface, and the underlying processing method has become "current data + research history". The company table saves the current version, and the task table saves the process and evidence chain.
+
+## Three triggering methods
+
+![](https://static-docs.nocobase.com/202607121807495.png)
+
+Background research shouldn’t just rely on a manual button. In real business, you may want to automatically complete the information after adding a new company, you may also need to make up historical records regularly, and you may also take the initiative to re-investigate before signing a contract or reviewing.
+
+The `New company background check` workflow handles automatic research after adding or updating a company. It listens to the data events of the company table and is triggered when the company name exists and the background information is empty. The AI ​​will not be called immediately after triggering, but will first check whether there are any unfinished tasks for the same company; if not, a new background check record will be created.
+
+![](https://static-docs.nocobase.com/202607121807374.png)
+
+The `Timing company background check` workflow is responsible for the continuous completion of historical data. It runs every 30 minutes, queries companies whose background information is still empty, and loops through batches. Inside the loop, we also check whether the task already exists, and then decide whether to create a new task. In this way, the scheduled task can be run repeatedly without creating multiple concurrently processed records due to repeated scanning.
+
+![](https://static-docs.nocobase.com/202607121807404.png)
+
+The `Manual company background check` workflow is bound to the "Run background check" button on the company details page, which is suitable for business personnel to proactively initiate a survey before visiting, signing a contract, or reviewing. Manual triggering and automatic triggering use the same set of follow-up links: a background check record is created first, and then the task execution workflow takes over the AI ​​investigation.
+
+![](https://static-docs.nocobase.com/202607121807793.png)
+
+These three entrances solve problems at different points in time, and are ultimately merged into the same background investigation record form. New triggers, scheduled triggers, and manual triggers are only responsible for recording the "need to investigate", and the specific execution, status management, and result writing back are handed over to subsequent workflows for unified processing.
+
+## Turn AI research into tasks
+
+`Do company background check` is the workflow that actually performs research. It listens for the `pending` record in the background check record table. Once the previous automatic, scheduled or manual process creates a task, this workflow will be triggered asynchronously.
+
+When executed, the workflow first queries whether the company still exists. If the company does not exist, the task will be closed and the description will be written; if the company exists, the task status will be switched to `processing`, and then the AI ​​employee will be called to generate the report. The prompt word of the AI ​​employee requires the output of two parts: a Markdown report that can be written directly into the company background field, and a summary for manual review.
+
+![](https://static-docs.nocobase.com/202607121808833.png)
+
+After AI returns structured results, the workflow first writes the report, summary, and old background content into the background check record, and then writes the new report back to the company record. This order avoids the problem of "only the latest results, no process records": the company page keeps the latest available content, and the task records retain the context before this generation and writing back.
+
+![](https://static-docs.nocobase.com/202607121808662.png)
+
+After tasking, batch processing will also become more natural. The scheduled workflow does not need to wait for each company's research to be completed, but is only responsible for creating multiple records to be processed; each record independently triggers the AI ​​survey. Multiple companies can advance in parallel, and if a certain task fails or times out, other companies will not be blocked.
+
+## Make AI results reviewable
+
+AI-generated reports are organized according to a fixed structure: company profile, core business, development history and capital background, market position and competitive perspective, sales follow-up judgment, and citation links. Business personnel can see not only the "conclusion", but also the risk tips and additional information given by AI in the summary.
+
+The background investigation record details page displays "Research report" and "Previous background" in tabs, and provides a "Copy" operation. In this way, you can quickly copy this report during discussion, review, or external communication, and you can also check changes against the old version.
+
+The record details also configure two AI worker tasks. in:
+
+- Improve the background research report: regenerate the report after adding information through dialogue, and write the results back to the company records
+- Compare the old and new background research reports: Read the old and new reports and let AI explain the substantial differences brought about by this update
+
+This allows AI not to stop at “generating text once” but to participate in the process of continuous maintenance, review and version comparison.
+
+![](https://static-docs.nocobase.com/202607121808444.png)
+
+## How to combine workflow
+
+Overall, this set of workflows can be divided into four layers.
+
+The first layer is responsible for creating tasks. `New company background check` is for newly added or updated companies, `Timing company background check` is for historical data completion, and `Manual company background check` is for manual initiative. They will all check whether there are any unfinished records before creating a task, reducing duplicate processing from the source.
+
+The second layer is responsible for performing tasks. `Do company background check` listens to the background check record, advances the pending task to processing, calls the AI ​​employee, and writes the report, summary, and company's current background fields upon completion.
+
+The third layer is responsible for providing controlled writeback capabilities to AI employees. As a tool-based workflow, `Update company background` restricts AI to only write specified records according to clear parameters to avoid over-exerting data modification permissions.
+
+The fourth layer is responsible for exception cleaning. `Clean overtime processing background check` runs every 30 minutes to clean up non-completed tasks that have not been completed for more than 15 minutes to avoid long-term processing of tasks after abnormal interruption.
+
+![](https://static-docs.nocobase.com/202607121808799.png)
+
+## What scenarios can be migrated to?
+
+What this scene shows is not an isolated form or a separate AI button, but a combination of several capabilities in NocoBase: the data table is responsible for carrying business objects and historical records, the page is responsible for viewing and triggering by business personnel, the workflow is responsible for scheduling and writing back, and the AI ​​staff is responsible for generating reviewable structured results.
+
+Similar models can be migrated to scenarios such as supplier admission, customer due diligence, contract risk preliminary review, lead quality scoring, public opinion tracking, and preliminary screening of investment and financing targets. As long as there are several requirements in the business such as "data needs to be continuously completed", "AI results need to be left behind" and "historical versions cannot be overwritten", a runnable, trackable and scalable automated process can be built in a similar way.
+
+## Reference documentation
+
+- [NocoBase Workflow](/workflow/)
+- [NocoBase AI Employee](/ai-employees/)
+- [NocoBase Workflow AI Employee Node ](/ai-employees/workflow/nodes/employee/configuration)
+- [NocoBase AI employee customization tool ](/ai-employees/features/tools)

package/dist/ai/docs/nocobase/ai-employees/workflow/nodes/knowledge/create-document.md

@@ -0,0 +1,79 @@
+---
+pkg: "@nocobase/plugin-ai-knowledge-base"
+title: "Workflow AI knowledge base node - Create document"
+description: "Use a collection event trigger and the AI knowledge base Create document node to synchronize newly added records to the AI knowledge base."
+keywords: "AI knowledge base,workflow,create document,collection event trigger,NocoBase"
+---
+
+# Create document
+
+<PluginInfo name="ai-knowledge-base"></PluginInfo>
+
+## Introduction
+
+In NocoBase, the **Create document** node writes new documents to an AI knowledge base from a workflow. In this example, it listens for new records in `Answers` and writes the answer body, document name, related questions, and a stable `Key` to the target knowledge base.
+
+AI knowledge base nodes are asynchronous nodes, so the workflow must use asynchronous execution. Prepare the target knowledge base and vector store first. See [Overview](./) for the full scenario.
+
+## Workflow structure
+
+The example workflow listens for new records in `Answers` and synchronizes each answer as one document in the target knowledge base.
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-create-editable-overview.png)
+
+Where:
+
+- `Collection event` listens for newly added collection records
+- `Create document` writes record content to the AI knowledge base
+
+## Before you start
+
+The example uses the `Answers` / `Questions` collections and target knowledge base described in [Overview](./). Before configuring the workflow, confirm that:
+
+- A Local knowledge base has been created, with file storage and vector store configured. See [Knowledge base overview](../../../knowledge-base/knowledge-base) and [Vector store](../../../knowledge-base/vector-store)
+- The `Answers.questions` relation field is ready. To review the collection structure, return to [Overview](./)
+- The workflow uses asynchronous execution. See [Workflow](../../../../workflow) for the basics
+
+## Configure the trigger
+
+Create an asynchronous workflow and select `Collection event` as the trigger. In the trigger configuration:
+
+- Set `Collection` to the collection to synchronize, such as `Main / Answers`
+- Set `Trigger on` to `After record added`
+- If the document needs fields from related collections, preload relation fields in `Preload associations`
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-create-trigger-config.png)
+
+The example preloads the `questions` relation field so related questions can be written to `Related questions`.
+
+## Configure the Create document node
+
+Add a `Create document` node after the trigger.
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-create-node-config.png)
+
+Key settings:
+
+These settings decide the knowledge base document name, body content, splitting behavior, and whether related questions are included.
+
+| Setting | Example value | Description |
+| --- | --- | --- |
+| `Knowledge base` | Target knowledge base | Select the same Local knowledge base prepared in the overview. The workflow creates a document in its `Documents` list and then enters splitting and vectorization. |
+| `Split document` | `No` | Controls whether the body is split into multiple segments. In this example, one answer is already a complete response. With splitting disabled, the whole document becomes one segment and works better with `Related questions`. |
+| `Related questions` | `Answers.questions.content` | Only takes effect when `Split document` is disabled. It writes common phrasings into `Related questions` for the current segment. It does not rewrite the body, but participates in retrieval. |
+| `Document type` | `Text` | Creates a knowledge base document from a text field. The example uses a collection field as the answer body, so `Text` is selected. Use `Attachment` if your business collection stores files or file URLs. |
+| `Content` | `Answers.Content` | Select the answer body as the knowledge base document body. The `Content` shown in segment management comes from this field. |
+| `Key` | `Answers.ID` | Sets the unique identifier of the knowledge base document. The example uses the `Answers` record ID so update and delete workflows can find the same document by the same `Key`. |
+| `Name` | `Answers.Title` | Sets the document name. The example uses the answer title, making it easier to identify the document in `Documents`, enter segment management, and run hit tests. |
+
+If `Related questions` references relation fields, configure `Preload associations` in the trigger first. Otherwise, the node can only access the current `Answers` record and cannot read the related `questions` data.
+
+:::warning Note
+
+`Key` is used by the later update and delete workflows. The Create, Update, and Delete workflows must use the same Key rule, usually the business collection record ID.
+
+:::
+
+## Next step
+
+The Create document workflow is the starting point of the synchronization chain. Continue with [Update document](./update-document) and [Delete document](./delete-document), then return to the "Verify the synchronization chain" section in [Overview](./) to verify add, update, and delete in order.

package/dist/ai/docs/nocobase/ai-employees/workflow/nodes/knowledge/delete-document.md

@@ -0,0 +1,71 @@
+---
+pkg: "@nocobase/plugin-ai-knowledge-base"
+title: "Workflow AI knowledge base node - Delete document"
+description: "Use a collection event trigger and the AI knowledge base Delete document node to delete knowledge base documents after collection records are deleted."
+keywords: "AI knowledge base,workflow,delete document,collection event trigger,NocoBase"
+---
+
+# Delete document
+
+<PluginInfo name="ai-knowledge-base"></PluginInfo>
+
+## Introduction
+
+In NocoBase, the **Delete document** node deletes a specified document from an AI knowledge base. In this example, it listens for deletion events in `Answers` and deletes the corresponding document from the target knowledge base by the same `Key` used by Create document, so removed data is no longer retrieved.
+
+The Delete document node is asynchronous. It depends on the `Key` written when the document was created, so it must use the same target knowledge base and the same `Key` rule as the Create and Update document nodes.
+
+## Workflow structure
+
+The example workflow listens for deletion events in `Answers` and deletes the document with the same `Key` from the target knowledge base.
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-delete-editable-overview.png)
+
+Where:
+
+- `Collection event` listens for collection deletion
+- `Delete document` deletes the document from the target knowledge base
+
+## Before you start
+
+The example uses the target knowledge base and `Key` rule from [Create document](./create-document). Before configuring the workflow, confirm that:
+
+- The Create document workflow is enabled and the target knowledge base contains documents with the same `Key`
+- The delete workflow can read this `Key` from trigger data. The example uses the `Answers` record ID
+- If you need to review the full synchronization chain, see the workflow responsibilities in [Overview](./)
+
+## Configure the trigger
+
+Select `Collection event` as the trigger. In the trigger configuration:
+
+- Set `Collection` to the collection to listen to, such as `Main / Answers`
+- Set `Trigger on` to `After record deleted`
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-delete-trigger-config.png)
+
+Deleting a document only needs the document `Key`, so relation fields usually do not need to be preloaded.
+
+## Configure the Delete document node
+
+Add a `Delete document` node after the trigger.
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-delete-node-config.png)
+
+Key settings:
+
+The Delete document node only needs to locate the document to delete. It does not need body content, title, or related questions.
+
+| Setting | Example value | Description |
+| --- | --- | --- |
+| `Knowledge base` | Target knowledge base | Select the same knowledge base used by the Create and Update document nodes. |
+| `Key` | `Answers.ID` | Select the exact same field used by the Create document node. The example continues to use the `Answers` record ID so the node deletes the document that corresponds to this answer. |
+
+:::warning Note
+
+The delete node does not look up documents by title or body content. It only locates documents by `Knowledge base` and `Key`. The Create, Update, and Delete workflows must use the same `Key`.
+
+:::
+
+## Next step
+
+After configuring the Delete document node, return to the "Verify the synchronization chain" section in [Overview](./) to verify add, update, and delete in order.

package/dist/ai/docs/nocobase/ai-employees/workflow/nodes/knowledge/index.md

@@ -0,0 +1,92 @@
+---
+pkg: "@nocobase/plugin-ai-knowledge-base"
+title: "Workflow AI knowledge base nodes overview"
+description: "Introduces the workflow AI knowledge base node scenario, collection structure, and how Create, Update, and Delete document nodes synchronize knowledge base content."
+keywords: "AI knowledge base,workflow,collection event trigger,knowledge base synchronization,NocoBase"
+---
+
+# Overview
+
+<PluginInfo name="ai-knowledge-base"></PluginInfo>
+
+## Introduction
+
+AI knowledge base nodes allow workflows to manage knowledge base documents directly. When business data changes, you can synchronize added, updated, and deleted records to the knowledge base, keeping knowledge base content aligned with your business collections.
+
+This documentation uses a Q&A scenario as an example: administrators maintain standard answers and common phrasings for each answer. Workflows synchronize the answer body and related questions to the same target knowledge base, so AI employees or other processes can later use those knowledge base documents.
+
+## Example scenario
+
+The example uses two business collections:
+
+- `Answers`: stores answer content that can be synchronized to the knowledge base
+- `Questions`: stores common phrasings related to each answer
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-overview-collections.png)
+
+`Answers` is the source collection for knowledge base documents. When an answer is added, a document is created. When an answer is changed, the document is updated. When an answer is deleted, the document is removed.
+
+`Questions` is not an independent source of knowledge base documents. It is attached to answers through the `Answers.questions` relation field and supplements the different phrasings that can point to the same answer.
+
+Before synchronization, prepare a Local knowledge base. You can name it according to your environment. The Create, Update, and Delete workflows only need to select the same knowledge base.
+
+## Answers collection structure
+
+`Answers` mainly uses three fields:
+
+- `Title`: used as the knowledge base document name
+- `Content`: used as the knowledge base document body
+- `questions`: relation to the `Questions` collection for related questions
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-overview-answers-fields.png)
+
+`questions` is a one-to-many relation field. Its source collection is `Answers`, target collection is `Questions`, both source key and target key use `ID`, and the foreign key is stored in `Questions`.
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-overview-answers-questions-relation.png)
+
+This relation field is preloaded when creating and updating documents. It writes related questions into `Related questions`, so one answer can be retrieved not only by body text but also by common phrasings.
+
+:::tip
+
+If your knowledge base documents do not need related questions, you do not need a relation field like `questions`. A stable `Key` field and body content are enough for create, update, and delete synchronization.
+
+:::
+
+## Workflow responsibilities
+
+The core of this example is three synchronization workflows. They process the same `Answers` data and write to the same target knowledge base:
+
+| Scenario | Trigger | Node | Purpose |
+| --- | --- | --- | --- |
+| Add answer | `Answers` collection `After record added` | `Create document` | Write the new answer to the AI knowledge base |
+| Update answer | `Answers` collection `After record updated` | `Update document` | Update the document by the same `Key` |
+| Delete answer | `Answers` collection `After record deleted` | `Delete document` | Delete the document by the same `Key` |
+
+`Key` is the most important field in the synchronization chain. Create, update, and delete workflows must use the same `Key` rule. The example directly uses the record ID of the `Answers` collection.
+
+:::tip Text content ingestion and vectorization
+
+When configuring a Create document or Update document node, if `Document type` is set to `Text`, the workflow first saves the text field selected in `Content` as a `.txt` document, then writes it to the target knowledge base. The knowledge base then generates segments according to the current segmentation settings, vectorizes enabled segments together with their related questions, and stores them in the vector database bound to the knowledge base.
+
+This process runs as an asynchronous workflow node. A completed node does not mean the document can be retrieved immediately. Usually, wait until `Status` in the `Documents` list changes to `Success` before checking segments or running hit tests.
+
+:::
+
+## Operation guides
+
+- [Create document](./create-document)
+- [Update document](./update-document)
+- [Delete document](./delete-document)
+- [Retrieve document](./retrieve-document)
+
+## Verify the synchronization chain
+
+After configuring the Create, Update, and Delete workflows, use the same `Answers` record to verify the full synchronization chain.
+
+After adding an answer, check in this order:
+
+1. Go to `Documents` in the target knowledge base, confirm that the new document appears, and check that `Status` is `Success`. See [knowledge base document management](../../../knowledge-base/knowledge-base/documents) for details.
+2. Click `Segments` for the document and confirm that `Related questions` includes common phrasings from `Answers.questions`. See [segment management](../../../knowledge-base/knowledge-base/segments) for details.
+3. Go to `Hit tests`, enter a keyword from the answer body or a related question, and confirm that the synchronized answer can be hit. See [hit tests](../../../knowledge-base/knowledge-base/hit-tests) for details.
+
+After the add synchronization is correct, update the same `Answers` record and confirm that the target knowledge base document changes. Finally, delete the record and confirm that the corresponding document is removed from the target knowledge base.

package/dist/ai/docs/nocobase/ai-employees/workflow/nodes/knowledge/retrieve-document.md

@@ -0,0 +1,87 @@
+---
+pkg: "@nocobase/plugin-ai-knowledge-base"
+title: "Workflow AI knowledge base node - Retrieve document"
+description: "Use the AI knowledge base Retrieve document node in a workflow to retrieve snippets and provide them as task context for an AI employee node."
+keywords: "AI knowledge base,workflow,retrieve document,AI employee,RAG,NocoBase"
+---
+
+# Retrieve document
+
+<PluginInfo name="ai-knowledge-base"></PluginInfo>
+
+## Introduction
+
+In NocoBase, the **Retrieve document** node searches an AI knowledge base in a workflow and returns matched document snippets. It is suitable before an AI employee node: retrieve knowledge base snippets with task input first, then pass the results to the AI employee as task context.
+
+The Retrieve document node is asynchronous. Before using it, configure the knowledge base and make sure it already contains vectorized documents.
+
+## Workflow structure
+
+The example workflow listens for new records in `Questions`, sends matching text to `Retrieve document`, and passes the retrieval result to a downstream `AI employee` node for task processing.
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-retrieve-overview.png)
+
+Where:
+
+- `Collection event` receives task input
+- `Retrieve document` retrieves snippets from the knowledge base
+- `AI employee` processes the task based on snippets
+
+In real business workflows, the result generated by the AI employee usually needs to be saved. You can continue with database operation nodes after it. This guide omits those later saving steps to focus on the Retrieve document node.
+
+## Before you start
+
+This example is an independent retrieval workflow. Before configuring it, confirm that:
+
+- The knowledge base already contains vectorized documents. For document maintenance, see [Knowledge base overview](../../../knowledge-base/knowledge-base) and [Create document](./create-document)
+- A data source for matching text is ready. The example uses the `Questions` collection
+- A usable downstream AI employee node is ready. See [AI employee node](../employee/configuration)
+
+## Configure the trigger
+
+Select `Collection event` as the trigger. In the trigger configuration:
+
+- Set `Collection` to the matching-text source collection, such as `Main / Questions`
+- Set `Trigger on` to `After record added`
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-retrieve-trigger-config.png)
+
+Each time an input record is added, the workflow automatically enters retrieval and task processing.
+
+## Configure the Retrieve document node
+
+Add a `Retrieve document` node after the trigger. It is recommended to change the node key to an easy-to-reference name, such as `retrieve_docs`.
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-retrieve-node-config.png)
+
+Key settings:
+
+- Select one or more knowledge bases in `Knowledge base`
+- Set `Input matching text` to the matching text field
+- Set `Top K` to the maximum number of snippets to return
+- Set `Score` to the minimum similarity threshold
+
+After execution, the node returns an array of document snippets. Each result contains `id`, `content`, `score`, and `metadata`, sorted by `score` from high to low.
+
+## Pass the result to the AI employee node
+
+In the `AI employee` node, write the retrieval result into the prompt. The example directly references the Retrieve document node output:
+
+```text
+Task input: {{$context.data.f_b2izf5j4xx0}}
+
+Knowledge base snippets:
+{{$jobsMapByNodeKey.retrieve_docs}}
+
+Use the snippets as context to complete the task. If the snippets are not enough, explain what information is missing.
+```
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-retrieve-ai-employee-config.png)
+
+`retrieve_docs` is the key of the retrieval node. If your node key is different, replace it with the actual key.
+
+:::tip
+
+If you only want to pass snippet text, reference the `content` field in the retrieval result through the variable selector. Passing the whole result array also includes `score` and `metadata`, which helps the AI judge source and relevance.
+
+:::

package/dist/ai/docs/nocobase/ai-employees/workflow/nodes/knowledge/update-document.md

@@ -0,0 +1,70 @@
+---
+pkg: "@nocobase/plugin-ai-knowledge-base"
+title: "Workflow AI knowledge base node - Update document"
+description: "Use a collection event trigger and the AI knowledge base Update document node to synchronize collection updates to the AI knowledge base."
+keywords: "AI knowledge base,workflow,update document,collection event trigger,NocoBase"
+---
+
+# Update document
+
+<PluginInfo name="ai-knowledge-base"></PluginInfo>
+
+## Introduction
+
+In NocoBase, the **Update document** node updates existing documents in an AI knowledge base from a workflow. In this example, it listens for updated records in `Answers`, finds the document in the target knowledge base by the same `Key` used by Create document, and rewrites it with the latest fields.
+
+The Update document node is asynchronous. It is not used as an independent entry point. It should follow the Create document workflow and use the same target knowledge base and the same `Key` rule as Create and Delete document nodes.
+
+## Workflow structure
+
+The example workflow listens for updated records in `Answers` and synchronizes the latest answer content to the same document in the target knowledge base.
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-update-editable-overview.png)
+
+Where:
+
+- `Collection event` listens for collection updates
+- `Update document` updates the document with the same `Key` in the target knowledge base
+
+## Before you start
+
+The example uses the target knowledge base and `Key` rule from [Create document](./create-document). Before configuring the workflow, confirm that:
+
+- The Create document workflow is enabled and has created at least one knowledge base document
+- `Key` uses a stable and unique business field. The example uses the `Answers` record ID
+- If updates should also synchronize related questions, preload the `questions` relation field in the trigger
+
+## Configure the trigger
+
+Select `Collection event` as the trigger. In the trigger configuration:
+
+- Set `Collection` to the collection to listen to, such as `Main / Answers`
+- Set `Trigger on` to `After record updated`
+- Use `Changed fields` to limit which field changes trigger the workflow, such as title or body changes
+- If the node needs relation fields, preload them in `Preload associations`
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-update-trigger-config.png)
+
+If `Changed fields` is not selected, any field change triggers this workflow.
+
+## Configure the Update document node
+
+Add an `Update document` node after the trigger.
+
+![](https://static-docs.nocobase.com/ai-employees/workflow/knowledge/2026-06-12/kb-update-node-config.png)
+
+Key settings:
+
+Most Update document settings are the same as [Create document](./create-document). Usually, continue using the same target knowledge base, splitting behavior, body field, and title field. If related questions also need to be synchronized, continue using `Answers.questions.content`.
+
+The important difference is `Key`: in the Update document node, `Key` is required and must be exactly the same as in the Create document node. The example continues to use `Answers.ID`, so the node can find and overwrite the same knowledge base document.
+
+:::warning Note
+
+If no document is found for the specified `Key`, the update fails. Before configuring this workflow, make sure the Create document workflow has already created documents with the same `Key`.
+
+:::
+
+## Next step
+
+After configuring the Update document node, continue with [Delete document](./delete-document). After all three workflows are configured, return to the "Verify the synchronization chain" section in [Overview](./) to verify add, update, and delete in order.