AI Colleagues and knowledge
Knowledge bases, tools, test cases, the chat and widgets.
Knowledge base
A knowledge base keeps what you know yourselves — manuals, terms, agreements — in a way Flow can search. Flow cuts every document into chunks and makes them searchable by meaning and by words. Agents and workflows then use the knowledge base as a source. Here you create one, put a document in it and try it out in the notebook: search and ask a question.
Step: Search knowledge base
With the step Search knowledge base a workflow fetches the best matching chunks from a knowledge base, by meaning and by words. The query may come from earlier steps, for example a customer's question. Later steps get a list of results: the text, the score and the document.
Step: Answer from knowledge base
The step Answer from knowledge base finds the best matching chunks and lets an AI model answer the question with them. The model only sees those chunks and cites them as sources: [1], [2]. That gives a workflow an answer you can check, for example to reply to a customer question by e-mail.
Step: Add document to knowledge base
You do not have to keep a knowledge base up to date by hand. The step Add document to knowledge base puts text or data from a workflow into a knowledge base as a document: in chunks, searchable and with metadata. That way every new note, e-mail or invoice ends up in the knowledge base by itself.
Step: Extract metadata
The step Extract metadata pulls typed data out of text, for example out of an invoice: the number, the dates, the amounts and the IBAN. Flow first uses fixed rules (labels, dates, amounts, IBAN with its check digits) and an AI model only when you set that up. Later steps know exactly what they get: the total is a number, the date a date.
Building block: HTTP call
An HTTP call is a building block that calls one API, for example the order status at a supplier. You make it once, test it and publish it; after that workflows use it as a step and agents as a tool. The API's key lives in a connection: the server adds it, the building block and the agent never see it.
Building block: AI prompt
An AI prompt is a building block with one AI call following a fixed template: summarise, classify, pull data out of text. You fix the instruction and the template; whoever uses it only gives the input. That way every workflow and every agent gets the same, tested prompt.
Building block: Extract text (regex)
Extract text pulls patterns out of text, such as order numbers, postcodes or IBANs, with a regular expression: without code and without AI, so fast and the same every time. This kind of building block comes from a plugin; plugins can add their own kinds to the toolbox this way.
Test cases for an agent
An agent that answers well today must do so tomorrow too — also after a new instruction or another model. Test cases record what the agent must do: an input, and what the answer must contain or which outcome you expect. Once an agent has test cases, you can only publish it when they all pass.
Step: AI agent
With the step AI agent a workflow hands one well-defined task to a published agent: write an e-mail, assess a message, make a summary. The agent works within its own guard rails and budget, and the workflow gets its answer back to carry on with. You choose which version the step uses and what the agent gets to see.
Buttons in a chat widget
A chat widget can also ask something: Approve or Reject, or a form to fill in. Every button starts a workflow, with the widget's data and the choice. That way a manager approves a purchase request without leaving the chat. A choice can be made only once, and the agent itself can never press a button.
Agent tool: knowledge base
Give an agent a knowledge base and it gets a search tool: it looks up the matching chunks itself when a question calls for it, and cites the source. It may only read the knowledge base. Documents set to Always include are given to it in full with every question.
Agent tool: workflow
An agent can start a workflow itself when it needs to look something up or do something. You choose which workflows it may use, say when, whether the workflow only reads or changes something, and how often per run. A workflow that changes something is only offered once the agent may change data.
Build with the AI Architect
Tell the AI Architect what you want. It works in small steps: it first reads Flow's guide and your workspace, writes down what it knows, and proposes the plan part by part: what it will make, in which order, and what stays for you. Only when you say ‘Build it’ does it build, part by part, with the same tools an AI app uses over MCP and never with more rights than you. You see every step live; a mistake Flow finds, it fixes itself; access it proposes instead of granting. Before it changes anything Flow records a version, so ‘Revert all’ puts everything back in one go.
Keeping an AI Colleague up to date together
Often two builders keep one AI Colleague up to date together: one sharpens the instructions, the other tests and publishes. Flow shows who has the AI Colleague open, says so at once when the other saves something, and never publishes a draft you have not seen: when your colleague changed it in the meantime, you first see their version and test it. Every version shows who published it. Below you see both builders' screens side by side.
Building block: Python code
Sometimes what you need does not fit a standard step: your own calculation, an unusual conversion. Then you write a Python building block: a function run(input, ctx) that returns a JSON value. It runs isolated, with a time and memory limit and no network by default. You test it with sample input and publish a fixed version.