Skip to main content

Knowledge Base

The Knowledge Base is where your documents live. Everything you upload here can be indexed for search, so your workflows and chatbots answer from your real content instead of guessing. Agents can also write back into it: meeting notes, lead context, and conversation summaries get filed as documents you can read and search later.

Open it from Knowledge Base in the left navigation. The page has two tabs: Knowledge Base (documents, covered on this page) and Collections (structured data tables, covered in Collections).

Knowledge bases​

A knowledge base is a named library of folders and documents inside one project. Most teams start with one base; you can add more to separate concerns, for example "Support docs" and "Sales docs".

Every project gets a default base automatically. It is labeled default in the base list and cannot be deleted. Documents that are not explicitly assigned to a base land in the default base.

Creating a base​

  1. Open the Knowledge bases side panel (on smaller screens, use the panel toggle labeled "Knowledge bases").
  2. Click New knowledge base.
  3. Enter a Name, for example "Sales docs".
  4. Choose Access: Public or Private.
  5. Click Create.

Access controls how the base handles documents:

AccessBehavior
PublicDocuments can live at the root of the base or in folders.
PrivateEvery document must live inside a folder, and each root folder is isolated by a tenant ID (see Visibility and tenant isolation).

Editing and deleting a base​

  • Click Edit base in the toolbar (or right-click anywhere and choose Edit knowledge base) to rename the base or change its access level.
  • To delete a base, select it in the side panel and click the trash icon. A base must be completely empty first: delete or move its documents and folders, then delete the base. The default base cannot be deleted.

The toolbar and breadcrumb​

The top of the page is a slim toolbar. On the left is a breadcrumb showing where you are: the base name, then each folder in the path, for example Sales docs > Contracts > 2026. The current location is shown in bold.

  • Click the base name to jump back to the root of the base.
  • Click any earlier folder in the breadcrumb to jump to that level.
  • You can also drop a document card onto a breadcrumb segment to move the document there (see Moving documents).

On the right are the page actions: New folder, Edit base, Crawl website, and Add document.

Folders​

Folders organize documents and, more importantly, define retrieval policy. The settings that matter (visibility, indexing, tenant) live on root folders, the folders at the top level of a base. Subfolders always inherit from their root folder.

Creating a folder​

  1. Navigate to where the folder should live (the root of the base, or inside another folder).
  2. Click New folder in the toolbar, or right-click and choose New folder.
  3. Enter a Folder name.
  4. For a root folder, also choose:
    • Visibility: Public or Private (tenant).
    • Indexing: Indexed (searchable by AI) or Not indexed.
    • Tenant ID (only shown for private folders, optional): an identifier for the customer or tenant this folder belongs to. Leave it empty to auto-generate one.
  5. Click Create folder.

When you create a folder inside another folder, the modal explains that the subfolder inherits visibility and indexing from its level-1 root folder, so those fields are not shown.

Each folder card shows two badges: Public or Private, and Indexed or Not indexed. Click a folder card to open it.

Folder settings​

Open Folder settings from the folder card's gear icon, or by right-clicking the folder. You can rename any folder. On a root folder you can also change visibility, indexing, and the tenant ID; the change cascades to every subfolder underneath it, and the documents inside are updated automatically:

  • Turning indexing off removes those documents from search (their stored vectors are deleted and their status returns to uploaded).
  • Turning indexing on, or changing visibility or tenant, re-queues the affected documents for indexing.

Deleting a folder​

Right-click the folder and choose Delete folder, or use the trash icon on the folder card. You will be asked to confirm: deleting a folder deletes the folder, all of its subfolders, and every document inside, including their indexed vectors and stored files.

danger

Folder deletion is recursive and permanent. There is no undo.

The right-click menu​

Right-clicking gives you quick actions anywhere on the page:

Where you right-clickActions
Empty spaceNew folder, Add document, Edit knowledge base
A folder cardThe above, plus Add document here, Folder settings, Delete folder
A document cardThe above, plus Download, Edit document, Delete document

Documents​

Uploading​

There are three ways to add documents:

  1. Click Add document in the toolbar. The upload modal shows the target ("Root" or the current folder name).
  2. Right-click a folder and choose Add document here.
  3. Drag files from your computer anywhere onto the page. An overlay appears ("Drop files to upload") and the files are attached to the folder you are currently viewing.

Supported file types include PDF, Word (.doc, .docx), PowerPoint (.ppt, .pptx), Excel (.xls, .xlsx), text (.txt, .md), data files (.json, .csv), and images (.png, .jpg, .jpeg, .gif, .webp).

note

In a private base you must upload into a folder. The toolbar's Add document button is disabled at the root, with the message "Private knowledge bases require uploading into a folder."

Crawling a website​

Your public website is usually the best first knowledge source, and the Crawl website button in the toolbar imports it without any copying and pasting:

  1. Click Crawl website and enter your website's URL.
  2. Pick a target folder, or let the crawl create one named after the domain.
  3. Set a page budget (up to 100 pages) and start.

The crawl runs in the background. The popup shows live progress — pages crawled, documents created, and the URL currently being fetched — and you can close it while the crawl continues; reopen it from the same button to check on it. Each crawled page becomes a markdown document carrying its source URL and is indexed automatically, just like an upload.

The crawler only follows pages on the same origin as the URL you enter, and only works on public websites.

Indexing status​

After upload, the platform parses the file, splits it into chunks, generates embeddings, and writes them into the vector store. Each document card shows a status badge:

StatusMeaning
uploadedStored, but not indexed (for example, its folder has indexing turned off).
indexingQueued or being processed.
indexedSearchable by agents.
failedParsing or indexing failed.

To retry a failed document, open Edit document and save it: a document in failed or uploaded status is re-queued for indexing when you save, as long as its folder allows indexing. Moving a document to a different folder or base also re-indexes it, because its retrieval scope changed.

Editing a document​

Click the gear icon on a document card (or right-click and choose Edit document) to change:

  • Name
  • Tags (comma separated, useful for filtering and maintenance)
  • Folder (moves the document; in a private base the document must belong to a folder)

Moving documents​

Document cards are draggable. Drop a card onto a folder card to move it into that folder, or onto a breadcrumb segment (including the base name) to move it up the tree. In private bases you cannot drop a document at the root.

Downloading and deleting​

  • Click the document name or the download icon to download the original file.
  • The trash icon (or right-click, Delete document) removes the document. The confirmation reads "Delete this document? This removes it from storage and search." Both the stored file and the indexed vectors are deleted.

How agents use the knowledge base​

Workflows and chatbots ground their answers through the built-in knowledge_base.search tool. It runs a vector-similarity search over the indexed documents and returns the best-matching chunks, which the agent quotes and reasons from. In the workflow designer you can attach it to an agent from the Tools & Connectors palette group as the Knowledge Base tool. If no base is specified, search falls back to the project's default base. See Workflows and Agents.

Choosing bases and folders in a chat​

In the Playground chat, click New chat to open the Set chat context modal. There you pick:

  • Workflow and Entry point: which assistant handles the chat.
  • Knowledge bases: add one or more bases the session may search. Leave empty to use the default base.
  • Private indexed folders: add specific private root folders to include. Public indexed root folders are included automatically.

The Session context panel in the chat sidebar shows the active workflow, entry point, knowledge bases, and folders for the current session, along with copyable developer IDs.

Agent-written notes​

Agents do not only read the knowledge base; they can file notes into it. A note is a markdown document (a meeting summary, lead context, a decision) written by an agent through the knowledge.note tool, stored as a real document, and indexed so later searches find it. Think of it as your workspace's shared memory.

The Vault node​

In the workflow designer, the Data & Memory palette group contains Vault (agent notes). Placing a Vault node on the canvas does two things:

  1. Every agent in the workflow gets the knowledge.note tool, so it can file notes on its own.
  2. Any step routed into the Vault node has its output stored as a note.

The node has one main setting, Vault folder name (default "Vault"). Notes are filed into a folder with that name in the knowledge base; the folder is created automatically on first use, public and indexed, so vault notes are retrievable through knowledge search. If a note does not name a base, it goes to the agent's configured base, falling back to the project's default base.

The meeting-vault recipe in the catalogue is a ready-made example: meeting notes posted to a workflow endpoint are structured by a librarian agent and filed into a "Meetings" vault folder.

The chat memory folder​

Each workflow has an optional Chat memory folder setting, in the workflow's settings modal on the Workflow page. Type a folder name (for example "Chat memory") and save. The folder is created in the project's default base if it does not exist. Leave the field empty to keep the Vault node's folder, or the built-in "Vault" folder.

With a chat memory folder set, the workflow's chatbot does two things:

  • Its agents get the knowledge.note tool automatically (even without a Vault node on the canvas) and file their notes into that folder.
  • When a conversation goes idle, the platform consolidates it and files a conversation abstract into the folder.

What conversation abstracts look like​

A conversation abstract is a small markdown note, titled "Conversation memory" followed by the channel, the conversation identifier, and the date. It contains:

  • A header line with the channel, the peer identifier, and the number of turns.
  • The time range the conversation covered.
  • A Summary section: two to four sentences on what the user wanted and the outcome.
  • A Durable facts section: bullet points worth remembering, such as preferences or unfinished business.

Because abstracts are indexed documents, both you and the agent benefit: you can open the folder and read what happened, and the agent finds the abstract again through knowledge_base.search the next time the same person writes in.

Visibility and tenant isolation​

Two folder settings control who and what can retrieve a document:

  • Indexing: documents under a root folder with indexing off are never searchable, by anyone. Use this for reference material you want stored but not fed to the AI.
  • Visibility: a public root folder is searchable by all requests in the project. A private root folder carries a tenant ID, and its documents are only retrievable when that tenant is in scope.

Tenant isolation matters most for the customer portal. Portal customers automatically get a tenant identity derived from their email address. When a customer chats or browses, they can only see public folders plus private folders whose tenant ID matches their own. Files a customer uploads through onboarding tasks land in their private folder, so one customer's documents never appear in another customer's answers.

All of this sits inside the platform's hard boundaries: knowledge is always scoped to your organization and project first, and folder and tenant filtering applies on top.