During the ConForAll campaign for constitutional reform and the 2026 referendum in Thailand, a vast number of questions surfaced daily, ranging from core principles to public participation. To address this, we developed an intelligent AI assistant for the campaign. Our goal was to ensure fast, reliable, and 24/7 access to campaign information for everyone.

Goal & Challenge

To develop an AI chatbot capable of providing instant FAQ responses 24 hours a day, the primary challenge was accuracy. Given that the topic involves sensitive political issues, the chatbot should not interpret or invent answers. Every response had to be strictly grounded in an official database, stay on point, and handle delicate content with high reliability.

Data Preparation & Ingestion

To ensure precision, we focused on the data pipeline from the start. Our data workflow is as follows.

FAQ data collection
Data cleaning
Embedding process (converting text into vectors)
Building vector database (storage for retrieval)

We utilized a "1 FAQ = 1 Vector Row" structure. This method allows the chatbot to pinpoint exact questions in the database more effectively than traditional methods that scan long documents. This ensures a higher match rate for both identical and semantically similar queries.

Why 1 FAQ = 1 Vector Row?

Most RAG systems use "chunking" or slicing long documents into pieces of a specific size but chunk boundaries do not always align with meaning. A question and its answer might end up in different chunks, forcing the chatbot to piece things together on its own.

For ConForAll, we set boundaries to match the project requirements. Each row embeds one FAQ, both question and answer concatenated together, such as "Q: When will the constitutional referendum happen? A: The referendum will take place at the same time as ...".

This one-to-one structure ensures the chatbot always retrieves a complete Q&A unit. There is no risk of pulling half an answer from a mid-sentence chunk, which provides incomplete context and is one of the main reasons some RAG systems hallucinate.

Design & Technology

To meet our strict accuracy requirements, we implemented the following solutions:

Retrieval-Augmented Generation (RAG): The AI searches the official database first and generates answers solely based on that data, rather than relying on its general pre-trained knowledge.
Vector Database: FAQ data is converted into embedding vectors and stored in a specialized database for high-speed retrieval.
Hybrid Search Implementation:
- Semantic Similarity: Understanding the intent behind a question, even if the user uses different phrasing.
- Full-text Search: Matching specific keywords and figures (such as voter statistics or article numbers) with 100% accuracy.
Strict Prompt Engineering: We designed a Strict Context Constraint prompt. The AI is forbidden from using external knowledge, improvising, or summarizing outside the provided context. This eliminates the risk of AI hallucination regarding political information.

Results

We successfully deployed a chatbot that delivers accurate and easy-to-understand information instantly. Being available 24/7 has significantly reduced the manual workload for campaign admins, who had to answer only questions unrelated to the prepared FAQ data. Furthermore, real-time usage data has provided valuable insights into what the public is most concerned about, allowing the team to improve communication strategies and campaign effectiveness for the future.

Key Highlights

Hallucination in RAG systems usually happens when the system pulls irrelevant context and lets the large language model connect the dots itself. For politically sensitive content where accuracy really matters, we built guardrails to address this.

Strict Context Constraint: The prompt forces the chatbot to answer only from the database. No outside knowledge allowed. If there's no matching information for a question, it returns a predefined fallback message instead of guessing.
1 FAQ = 1 Vector Row: The data structure itself reduces the chance of the chatbot pulling the wrong context, which directly improves answer accuracy.
Hybrid Search: By combining semantic and full-text search, we solved the common AI struggle with both intent and specifics. Semantic search catches questions that are phrased differently but mean the same thing, while full-text search catches specific terms like law names or numbers that semantic search might miss.

ConForAll Chatbot

Related Project

Chatbot for Office of International Affairs and Global Network, Chulalongkorn University

News Chatbot for Thai PBS