Build a Multi-Model Workflow

Create multi-stage LLM pipelines using the Visual Workflow Builder or API. Chain models sequentially, run them in parallel, and merge outputs.

Prerequisites

• A Xilos account with admin access
• Multiple LLMs configured in Model Config
• An understanding of your pipeline goals

Workflow Concepts

A workflow consists of:

• Stages — Each stage uses one LLM to process input and produce output
• Execution Mode — sequential (chain) or parallel (run simultaneously and merge)
• Merge LLM — For parallel mode, an LLM that combines outputs from all stages
• Merge Prompt — Instructions for how the merge LLM should combine results

Step 1: Plan Your Pipeline

Decide what you want the workflow to accomplish:

• Sequential example: Model A summarizes → Model B translates → Model C formats
• Parallel example: Model A analyzes code, Model B checks security → Merge LLM combines both analyses

Step 2: Open the Visual Workflow Builder

1. Navigate to Workflow Builder in the sidebar.
2. You'll see a canvas with a left sidebar of draggable node types:
   • LLM Stage — A model processing step
   • Merge — Combines multiple outputs
   • Conditional — Branch based on output
   • Input — Starting point
   • Output — Ending point

Step 3: Build Your Workflow

1. Drag an Input node onto the canvas.
2. Drag an LLM Stage node and connect it to the Input.
3. Double-click the LLM Stage to edit properties:
   • Select the model from the dropdown
   • Write a system prompt for this stage
   • Set the temperature slider
4. For sequential chains, drag another LLM Stage and connect it.
5. For parallel execution, drag multiple LLM Stages from the same source.
6. For parallel, drag a Merge node and connect all parallel outputs to it.
7. Drag an Output node and connect it as the final step.

Step 4: Configure Merge Settings (Parallel Only)

If using parallel execution:

1. Double-click the Merge node.
2. Select the merge LLM (the model that will combine outputs).
3. Write a merge prompt — instructions for how to combine the parallel outputs.

Example merge prompt:

"Combine the following analyses into a single comprehensive report. Maintain all key findings from each analysis and organize by theme."

Step 5: Test the Workflow

1. Click the Test Run button.
2. Enter a test query.
3. Review the output from each stage.
4. Adjust system prompts and connections as needed.

Step 6: Save the Workflow

1. Click Save.
2. Enter a name and description.
3. The workflow is serialized into JSON format and stored.

Workflow JSON Format

The workflow is stored as:

{
  "stages": [
    {
      "name": "stage1",
      "stage_type": "llm",
      "llm_id": "uuid",
      "system_prompt_override": "You are a code analyzer.",
      "temperature": 0.7,
      "input_source": "previous_output"
    }
  ],
  "execution_mode": "sequential",
  "merge_llm_id": null,
  "merge_prompt": null
}

Using Workflows via API

Workflows can also be created and managed via the API:

# Create a workflow
curl https://api.xilos.ai/api/v1/workflows \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_XILOS_API_KEY" \
  -d '{
    "name": "Code Analysis Pipeline",
    "description": "Analyze code and check security",
    "execution_mode": "parallel",
    "merge_llm_id": "uuid-of-merge-model",
    "merge_prompt": "Combine both analyses into a report.",
    "stages": [...]
  }'

See the Workflows API for the full reference.

Best Practices

• Start simple — test with a 2-stage sequential workflow first
• Write clear, specific system prompts for each stage
• Use parallel execution when stages are independent
• Use the Merge LLM to synthesize, not just concatenate
• Test with real queries before saving
• Monitor workflow execution in the Query Log