Subgraph Node

Subgraph Node

GraSP supports modular graph design through the subgraph node. Subgraphs allow you to encapsulate and reuse a sequence of nodes and edges as a single unit within a main graph. This makes complex workflows more organized, reusable, and easier to maintain.

You can include a subgraph in your main graph and customize the configuration of its internal nodes as needed for your specific workflow.

Example Configuration

Subgraph Definition (generate_question subgraph)

# Example: tasks/examples/custom_subgraphs/generate_question_subgraph/graph_config.yaml
graph_config:
  nodes:
    persona_sampler:
      node_type: weighted_sampler
      attributes:
        num_turns:
          values: [2, 3, 4, 5]
        tone1:
          values: [professional, casual, friendly]
        persona1:
          values: [high school teacher, software engineer]

    paraphrase_question:
      node_type: llm
      output_keys: text
      prompt:
        - system: |
            Assume you are {persona1} persona.
            Paraphrase the question while preserving all details.
        - user: |
            QUESTION: {p}
      model:
        name: gpt4o
        parameters:
          temperature: 0.1

  edges:
    - from: START
      to: persona_sampler
    - from: persona_sampler
      to: paraphrase_question
    - from: paraphrase_question
      to: END

Main Graph Including the Subgraph

nodes:
  generate_question:
    node_type: subgraph
    subgraph: tasks.examples.custom_subgraphs.generate_question_subgraph
    node_config_map:
      paraphrase_question:
        output_keys: rephrased_text
        prompt_placeholder_map:
          p: prompt
        model:
          name: gpt4o
          parameters:
            temperature: 1.0

Configuration Fields

• node_type:
  Should be set to subgraph to indicate this node is a subgraph reference.

• subgraph:
  The import path or module reference to the subgraph definition. This should point to the graph configuration file that defines the subgraph.

• node_config_map(Optional) :
  Dictionary to customize nodes inside the subgraph for this usage.

• The keys are node names within the subgraph (e.g., paraphrase_question).
• The values are configuration fields you wish to modify for that specific subgraph node, such as output_keys, model, prompt, etc.

• Note: Only the fields explicitly provided in node_config_map are overridden for the specified subgraph node. All other settings for that node remain as defined in the original subgraph.

  Supported fields include (but are not limited to): - output_keys: Override the output keys for the subgraph node. - prompt_placeholder_map: Remap input placeholders in prompts. For example, if the subgraph uses {p} but your main graph uses prompt, you can set prompt_placeholder_map: { p: prompt }. - model: Override model name or parameters. - Any other valid node configuration field.

Example: Using prompt_placeholder_map

If a subgraph node’s prompt uses p but your main graph contains prompt as input state variable, you can map this as follows:

prompt_placeholder_map:
  p: prompt

This will update all occurrences of {p} in the subgraph prompt of that node to {prompt}.

Notes

• Customization: With node_config_map, you can adapt subgraph nodes to fit your main graph’s needs—change output keys, update model parameters, or map input variables as needed. Only the fields you specify are changed; all others remain as originally set in the subgraph.
• Prompt Placeholder Mapping: Use prompt_placeholder_map to easily align variable names between your main graph and subgraph's prompt, without changing the original subgraph.
• Reusability: Subgraphs help you avoid duplication by making it easy to reuse common sequences or workflows across multiple graphs.
• Compatibility: Adding subgraph nodes does not affect existing graphs that don’t use this feature.

---

Tip: Use subgraphs to modularize common tasks such as data preprocessing, prompt engineering, or multi-step LLM workflows, and include them in different main graphs with custom settings as needed.

---