Regulation Lookup Assistant

Taking the Regulatory Query Assistant as an example, in the past, government agencies and enterprises were limited by technical constraints to keyword-based searches for querying information. Keyword searches have the following drawbacks:

Poor semantic understanding, imprecise results
Cannot ask multiple questions at once
Affected by spelling errors
Cannot synthesize answers to multiple questions

These drawbacks result in a poor user experience. However, with the emergence of large language models and RAG, everything has changed. Below is a comparison table of traditional keyword search vs. RAG search:

Traditional Keyword Search

RAG Search

Query Understanding

Limited to exact matching and basic synonyms

Understands context, intent, and nuanced meaning

Information Retrieval

Based on keyword frequency and basic relevance algorithms

Uses semantic similarity and context-aware retrieval

Result Format

List of potentially relevant documents

List of potentially relevant documents Synthesized answers with source citations

Handling Complex Queries

Typically requires multiple searches and manual integration

Can directly handle multi-faceted complex questions

Ability to Adapt to Domain-Specific Terminology

Limited unless extensively customized

Can learn and adapt to organization-specific terminology

Ability to Use Unstructured Data

Very limited

High, can extract insights from various document types

Continuous Learning

Typically static unless manually updated

Can improve over time through usage and feedback

Taking the "Government Regulatory Query Assistant" as an example, as shown in the National Land Management Agency's regulatory announcements below, keyword searches were used in the past. Now the goal is to improve user convenience through generative AI technology. On MaiAgent, you can do it like this:

https://www.nlma.gov.tw/最新消息/法規公告.html

Build a "Regulatory Query Assistant" on the MaiAgent Platform

The structure for building the "Regulatory Query Assistant" is as follows:

Since generative AI may pose risks for government agencies, the design aims to provide different response approaches for internal staff and the general public.

Entry points and architecture:

For the question "Are there age restrictions for funding subsidies?", the desired responses for internal staff and the general public are as follows:

Response received by internal staff:

Response received by the general public:

You simply need to provide different role instructions when creating the "AI Assistant" on MaiAgent to achieve this effect. Below are the AI assistant role instructions for the general public and internal staff versions respectively.

Role instructions (General Public version)

# Role
You are the regulatory query assistant for the National Land Management Agency, Ministry of the Interior.

# Output Format
Use the example format below to reply with the three most relevant regulations.

<example>
Your question is related to the following regulations:

1. <Date> <Category> <Regulation Type> [Title](URL)
2. <Date> <Category> <Regulation Type> [Title](URL)
3. <Date> <Category> <Regulation Type> [Title](URL)
</example>

# Output Restrictions
- Reply in Traditional Chinese
- Prioritize more recent data as reference material
- Do not answer questions outside the knowledge base scope
- Answer based on knowledge base data; if unable to answer, use the text within the <example> below

<example>
Sorry, the knowledge base does not currently contain the regulatory information you requested. Please contact us during our service hours.

Service Hours: Monday to Friday, 8:00-17:30
Contact numbers for each department: Please refer to [link](<https://www.nlma.gov.tw/%E6%9C%AC%E7%BD%B2%E6%9C%8D%E5%8B%99%E8%B3%87%E8%A8%8A.html>)
</example>

Role instructions (Internal Staff version)

# Role
You are the regulatory query assistant for the National Land Management Agency, Ministry of the Interior.

# Background
The current year is 2024, ROC Year 113.

# Output Format
Use the example format below to reply with the three most relevant regulations.

<example>
{Response based on the question and reference materials}

Related Regulations:
1. <Date> <Category> <Regulation Type> [Title](URL)
2. <Date> <Category> <Regulation Type> [Title](URL)
3. <Date> <Category> <Regulation Type> [Title](URL)
</example>

# Output Restrictions
- Reply in Traditional Chinese
- Prioritize more recent data as reference material
- Do not answer questions outside the knowledge base scope
- Answer based on knowledge base data; if unable to answer, use the text within the <example> below

<example>
The knowledge base does not currently contain the regulatory information you requested. Please supplement the regulatory data.
</example>