Doc-Driven Integration Guide (DDI)

Overview

Doc-Driven Integration (DDI) treats documentation as a contract. By following a unified documentation specification (DocSpec) and plan declaration, new DSLs, frontends, and backends can be integrated without modifying the AIKG core, reducing coupling and maintenance costs.

Orchestration Plan Integration

Enable and configure through the docs_dir field in the Task Orchestration Plan. Examples: aikg/python/ai_kernel_generator/config/vllm_triton_ascend_coderonly_config.yaml (Ascend) or aikg/python/ai_kernel_generator/config/vllm_triton_cuda_coderonly_config.yaml (CUDA):

Basic Configuration Structure

docs_dir:
  designer: "path/to/designer/docs"  # Designer reference document directory
  coder: "path/to/coder/docs"        # Coder reference document directory

DocSpec Requirements

Based on code analysis, the following documents are required for proper Agent operation:

Required

• basic_docs.md - DSL basic documentation

Optional (Recommended)

• basic_docs.md - DSL basic documentation
• api/api.md - API interface documentation
• suggestion_docs.md - Expert suggestion documentation
• examples/ directory - Contains example files starting with framework names (e.g., mindspore_*.py)

Usage

1. Prepare the directory structure:

your_docs_dir/
├── basic_docs.md
├── api/
│   └── api.md  
├── suggestion_docs.md
└── examples/
    ├── mindspore_example.py
    ├── torch_example.py
    └── ...

2. Specify paths in the plan file:

docs_dir:
  designer: "path/to/your_docs_dir"
  coder: "path/to/your_docs_dir"

3. Apply the configuration:

from ai_kernel_generator.config.config_validator import load_config

config = load_config(config_path="./your_custom_plan.yaml")
task = Task(op_name="custom_op", task_desc="...", config=config)

By following DocSpec and declaring docs_dir in the plan, you can efficiently integrate new DSLs/frontends/backends into AIKG.