Output Record Generator

Output Record Generator

The Output Record Generator is a flexible component that translates the final state of the graph into the desired output format for each record.

Key Concept:

Configure the final record fields in a YAML block (called output_map), telling the generator how to map fields from the graph's state or use static values—and optionally apply transforms.

Usage:

In the graph_config.yaml, declare an output_config section specifying a generator, such as processors.output_record_generator.CodeGenOutputGenerator

1. Simple Derived Class

Subclass the BaseOutputGenerator and implement the generate() method that returns the final record from the state.

Example:

• YAML Configuration

  output_config:
    generator: grasp.processors.output_record_generator.CodeGenOutputGenerator
  

• Python Code

  class CodeGenOutputGenerator(BaseOutputGenerator):
      def generate(self, state: GraphState) -> dict[str, Any]:
        if "messages" not in state:
            return None
  
        chat_format_messages = utils.convert_messages_from_langchain_to_chat_format(
            state["messages"]
        )
        if (
            len(chat_format_messages) < 1
            or "no more feedback"
            not in chat_format_messages[-1]["content"].lower().strip()
        ):
            return None
        # remove the last message if it contains "no more feedback"
        chat_format_messages = chat_format_messages[:-1]
        chat_format_messages.insert(
            0,
            {
                "role": "user",
                "content": state["rephrased_text"].replace(
                    "PARAPHRASED QUESTION: ", ""
                ),
            },
        )
        return {
            "id": state.get("id", ""),
            "conversation": chat_format_messages,
            "taxonomy": [{"category": "Coding", "subcategory": ""}],
            "annotation_type": ["mistral-large"],
            "language": ["en"],
            "tags": ["mbpp", "reannotate", "self-critique"],
        }
  

2. YAML driven Mapping & Transform

In this approach, we define a YAML configuration that maps fields from the graph's state or uses static values—and optionally apply transforms.

Example:

• YAML Configuration

output_config:
  generator: grasp.processors.output_record_generator.CodeGenOutputGenerator

  output_map:
    id:
      from: "id"  # Copy from state["id"]
    conversation:
      from: "messages"
      transform: "build_conversation"  # Apply a transform method on state["messages"]
    taxonomy:
      value:
        - category: "Coding"  # Hard-coded literal value
          subcategory: ""
    annotation_type:
      value: [ "mistral-large" ]   # Hard-coded literal list
    language:
      value: "en"   # Hard-coded literal value
    tags:
      value: ["mbpp", "reannotate", "self-critique"]   # Hard-coded literal list

• Python Code

  class CodeGenOutputGenerator(BaseOutputGenerator):
      """
      Example specialized generator class, which defines a transform method
      for building a conversation from messages, removing 'no more feedback', etc.
      """
  
      @staticmethod
      def build_conversation(data: Any, state: dict[str, Any]) -> Any:
          chat_format_messages = utils.convert_messages_from_langchain_to_chat_format(data)
  
          # Example logic:
          if (
              not chat_format_messages
              or "no more feedback" not in chat_format_messages[-1]["content"].lower().strip()
          ):
              return None
  
          # remove the last message
          chat_format_messages.pop()
  
          # insert rephrased question if available
          if "rephrased_text" in state and state["rephrased_text"]:
              question = state["rephrased_text"].replace("PARAPHRASED QUESTION: ", "")
              chat_format_messages.insert(0, {"role": "user", "content": question})
  
          return chat_format_messages
  

• generator: The dotted Python path to the output generator class (e.g., CodeGenOutputGenerator).

• output_map: Each key (e.g., id, conversation, tags) will become a field in the final record.
• from: Means “read from the in-memory state with that key.”
• value: Means “use this literal value.”
• transform: Means “call a method on the generator class to transform the data.” This is optional.

Note: If we include "transform": "my_custom_logic" in the YAML, you then a method named my_custom_logic must be defined in the generator class.

Here, we define a build_conversation method that takes the data and state as arguments and returns the transformed conversation.