How workflows help in everyday use
The same blocks can produce very different experiences. These three scenarios first show the Semantics Chat experience, then the matching Studio architecture.
Internal FAQ chatbot
User: How many days can I work from home each week?
HR assistant: You can work remotely up to three days a week. Submit the request in the HR portal no later than the previous day.
Behind the scenes, the instruction (prompt) describes the policy, scope, and tone. Each message sends the instruction, conversation history, and question to the model.
For this case without tools, disable Tools in the LLM instruction and request. Quality mainly depends on a precise instruction, especially on what the assistant must refuse or acknowledge it does not know.
RAG document search
User: What is the travel expense reimbursement process?
Document assistant: According to the 2026 expense policy, complete the form within 30 days and attach the original receipts.
Before answering, the model uses MCP Retriever to query your data sources. The most relevant retrieved passages return in the current turn’s history, and the model writes an answer grounded in those sources.
When the chosen block offers a source connection, use Data Sources to choose the embedding model and useful content. See source and tool connections and the complete RAG workflow.
Assistant with MCP tools
User: Generate an image of a cat.
Assistant: All right, I’ll generate the image.
The model receives the tool catalog and decides whether an action is required. Studio runs the requested call. The result goes back to the model to be rephrased and returned to the user.
Without a tool, the assistant can only describe a process. With MCP, it can act. See the complete MCP agent.
Choose an architecture
| Need | Recommended architecture |
|---|---|
| Answer within a stable scope described by an instruction | Simple FAQ |
| Answer from internal documents | RAG through MCP Retriever |
| Run an action or query a service | MCP |
| Combine internal documents with external data or actions | MCP Retriever + other MCP tools |
| Remember information across messages | CONVERSATION variables + conditions |