An AI Colleague with knowledge and tools

In the portal, under Environments, click Development, and there Open. Flow opens in a new tab. You build everything here first, in development; colleagues only notice once you publish.

The chat is where colleagues talk to your AI Colleagues. It is off in every environment until an administrator switches it on: open Administration in the bar on the left, choose Chat with AI Colleagues and switch it on. Below it you decide which data the chat may reach.

A knowledge base is what the AI Colleague knows. Click ＋ next to General and choose New knowledge base. Under Settings you give it a name and a description; the AI Colleague reads that description to decide when to search it. Under Embedder you choose what Flow turns text into vectors with: an embedding model searches by meaning, Built in works without internet and finds shared words and parts of words; Flow already picks the best one your environment has. Under Cutting into chunks you choose By heading: every chapter of the course material becomes one piece, so a search finds exactly the right chapter.

Under Documents, click Add document, choose File and the course material (Markdown; Word, PDF and Excel work too). Flow cuts it into pieces, by heading, and turns every piece into a vector: a series of numbers that captures its meaning. That way a question finds the right piece even when it uses other words than the text. Next to the document you see how many pieces it became.

What an AI Colleague looks up or changes lives in an ordinary Flow table. You already have the students in a list: click ＋ next to General, choose Table from a file and choose the file. Flow recognises the columns name, module and status.

Call the table Students. Module and status have only a few different values; Flow turns them into a fixed list, so no typos get in later, not even when the AI Colleague adds a student. Click through to Import.

An AI Colleague never just touches your data: it may only start the workflows you give it, and those do exactly what you built. Create a new workflow Look up student. In the trigger's sample data you put what it receives: a name. Those fields become the input the AI Colleague fills in later.

Add the Table step and choose Students; What does this step do is already set to Look up rows. With ＋ Condition you say which rows: where the name equals {{name}}, the name from the trigger. Connect the trigger to the step, save and run: at the bottom you find the student from the example.

The second workflow, Add student, also receives a name. Its Table step does Add one row: the name from the trigger, Module 1 and the status enrolled. That way you decide how a new student gets in, not the AI. Save it; there is no need to run it, the AI Colleague will do that.

Click ＋ next to General and choose New agent. Under Settings you give it a name, switch on Available in the chat and write a greeting: that is the first thing a student sees.

Under Brain you choose the AI model. A model that is not set up is greyed out; ask your administrator. The goal and the instructions say what it does, and what never: that keeps it on the course and stops it making things up.

Click the knowledge base: that is where it searches, and it names the source with every answer. Tick both workflows and say under When to use it? what they are for. Flow sees for itself that Look up student only reads and Add student changes something. An AI Colleague may only change things once you switch on May change data under Guard; until then it is not even offered that workflow.

Under Try you give a question and click ▶ Try. Under What happened you see that it searched the knowledge base first and only then answered: that is RAG in a nutshell.

Ask it to add a student. It picks the workflow Add student itself and fills in the name. In a test, Flow does not really carry out a change but simulates it: you see what it would do, without a row being added to your table.

After a passing test you publish. Colleagues in the chat always talk to the published version; what you change after that stays a draft until you test and publish again.

Open the chat with the chat icon on the left and pick your AI Colleague. Ask a question about the course. While it works you see it searching the knowledge base; below the answer it says which document it came from. Click Sources for the piece of text itself: that way anyone can check it is right.

Start a conversation about something else with New conversation and ask it to look up a student. Now it does not search the knowledge base but starts the workflow Look up student with the name as input, and answers with what is in the table.

In a new conversation you ask it to add a student. It starts Add student; because it may change data and this is not a test, it really happens. Every answer is recorded as a run, with every step, so you can always see what it did.

Open the table Students in the explorer: the new student is there, in module 1, enrolled — exactly as your workflow decides.