GenAI (源内) Generative AI Utilization Platform

The e-Laws Data Update Workflows are automated GitHub Actions pipelines designed to periodically fetch the latest legal XML data from the e-Gov Laws API and reflect it in the BigQuery e_laws_search dataset.

Processing Flow:

1. Download latest legal XML data from e-Gov Laws API.
2. Unzip the downloaded files.
3. Create a date-stamped GCS bucket (e.g., {PROJECT_ID}-elaws-YYYYMMDD) with a 30-day automatic deletion lifecycle policy.
4. Execute run_entire_pipeline.sh to load data into BigQuery:
   • Source Layer: Overwrites the source_laws_latest table (fixed name).
   • DWH & App Layers: Automatically updated.
5. Clean up temporary files.

GenAI (源内) is a generative AI utilization platform developed and operated by the Digital Agency of Japan. It provides an environment for government officials to use task-specific generative AI applications quickly, safely, and easily.

The ecosystem consists of two main components:

1. GenAI Web (源内 Web): The web application that users interact with directly.
2. Administrative AI Apps (行政実務用 AI アプリ): Microservices that leverage generative AI.

This repository contains implementations of administrative AI apps used in central government ministries. These apps can be built in environments independent of GenAI Web as long as they comply with the protocol used between GenAI Web and the microservices. Once built, they can be registered to GenAI via GUI operations.

GENAI (Gennai) is a generative AI platform developed by the Digital Agency, Government of Japan, designed to provide secure and fast task-specific AI applications for government officials.

The platform architecture consists of two primary components:

• GENAI Web: The web application interface used by end users.
• AI Applications for Government Operations: Microservices powered by generative AI that handle specific tasks.

This repository contains a subset of the AI applications deployed in central government ministries. Developers can build AI applications in environments independent of GENAI Web, provided they comply with the communication protocol between the application and GENAI Web. Once compliant, applications can be registered to GENAI via GUI operations.

This API enables data analysis using Azure OpenAI's Code Interpreter capabilities via Azure Functions.

Key Features:

• Upload Excel files (.xlsx, .csv).
• Perform data analysis using Azure OpenAI Code Interpreter.
• Automatic generation of graphs (with support for Japanese fonts).
• Returns both analysis results and generated graphs.

Architecture: Client $\rightarrow$ Azure Functions $\rightarrow$ Azure OpenAI (Code Interpreter)

Azure API Management (APIM) policies allow you to transform requests and responses between a simplified client-facing API and a complex backend API (such as vLLM's OpenAI-compatible API).

This is useful for:

• Simplifying Requests: Hiding model-specific prompt formats or special tokens from the client.
• Reducing Response Verbosity: Stripping away unnecessary metadata so the client only receives essential information.
• Improving Usability: Providing a clean, purpose-built API specification for end-users.

Because Japanese legislation often contains text describing prohibited or discriminatory acts, standard AI safety filters (which might block such content) are set to BLOCK_NONE for the Gemini API.

To prevent misuse and ensure quality, the system implements guardrails through System Prompts (defined in modules/api/functions/src/prompts.py) rather than API-level safety settings. These prompts enforce:

• Role Limitation: Restricting the model to a specialized legal query responder.
• Accuracy: Instructing the model not to speculate on uncertain content.
• Mandatory Citations: Requiring structured output with article numbers and source links.

To ensure type safety and IDE autocompletion, use config_types.py to define the structure of your configuration items using Python dataclasses and Enum. This prevents errors caused by typos in dictionary keys and provides explicit type validation for configuration parameters.

from dataclasses import dataclass
from enum import Enum

class ModelProvider(Enum):
    ANTHROPIC = "anthropic"
    AMAZON = "amazon"

@dataclass
class AnswerGenerationConfig:
    model_id: str
    temperature: float
    max_tokens: int
    system_prompt: str

The system manages legal data using a three-layer structure:

1. Source Layer: Raw legal XML data.
2. DWH (Data Warehouse) Layer: Historical data containing all versions.
3. App Layer: The latest data optimized for the API, including vector indexes for fast retrieval.

Understanding the distinction between utilities and services helps maintain a clean architecture:

Example:

• format_timestamp(timestamp) $\rightarrow$ utils/ (Generic, stateless)
• track_bedrock_usage(model_id, tokens) $\rightarrow$ services/ (Bedrock-specific, stateful)