AlgoVote - RAG-based Policy Comparison Chatbot

"An AI system that compares candidates’ policies side-by-side and answers voter questions with verified information."

Problem – Voters struggle to compare policies objectively

Although voters have access to candidate information, meaningful policy-based comparison is extremely difficult due to:

• Long and complex policy documents

• Different writing styles and terminology across candidates

• Media content that is often biased or incomplete

• Lack of structured tools to highlight actual policy differences

As a result, voters face an environment where information is abundant, but comparison and evaluation are hard.

Insight – Voters don’t want more information; they want clearer comparisons

User interviews and testing revealed several recurring needs:

1. Not just summaries — direct, structured comparisons between candidates

2. A Q&A interface that understands and responds to policy-specific questions

3. A filter to determine whether a question is actually about policy

4. Natural follow-up questions that stay within context

Voters needed an interactive, policy-grounded comparison tool, not another information portal.

Solution – Algovote RAG Policy Agent

A vector-search and agent-based system that compares candidates’ policies and answers voter questions accurately.

Key Features

1) Policy Question Filtering

• Determines whether the user’s query is related to policy

• Prevents hallucinations by separating non-policy questions

2) Intent → Slot Extraction

• Extracts key parameters like policy domain, comparison target, and candidate names

• Example: “Whose youth housing policy is more realistic?”

3) Candidate-Specific RAG Search

• Vectorized embeddings of each candidate’s policy documents

• Retrieves the most relevant segments by topic/section

4) Direct Candidate Comparison

• Supports comparing two or more candidates

• Highlights differences in approach, feasibility, beneficiaries, budget scope, etc.

5) Follow-up Rewrite Chain

• Normalizes ambiguous follow-up questions

• Ensures consistent context across long conversations

6) Missing-Information Fallback

• Detects when the policy documents lack answers

• Responds transparently without fabricating information

User Flow

User Question
→ Policy Question Filter
→ Intent & Slot Extraction
→ Policy Document Vector Search
→ Candidate Comparison Generation
→ Follow-up Question Rewrite

My Role – Product Designer & AI Agent Builder

Algovote was fully designed and built by me, reaching DAU 1,000 during peak usage.

1) Product Design & Information Architecture

• Organized policy documents by domain

• Standardized structure for inconsistent public policy formats

• Designed UX patterns suited for election-related decision-making

2) RAG Architecture Design

• Created candidate-specific embedding tables

• Designed chunking strategy optimized for policy documents

• Built exception dictionary for comparison-based questions

3) AI Agent Development

• Implemented policy-filtering chain (filter_prompt)

• Built follow-up rewrite chain

• Designed routing logic for candidate comparison queries

4) Custom Memory Management

• Implemented custom ConversationBufferMemory

• Maintained accuracy and context in long multi-turn dialogues

Impact – Real user validation and strong public engagement

• Achieved 1,000+ DAU shortly after launch

• Gained organic traction on SNS and community platforms

• Improved policy matching accuracy using real user queries

• Demonstrated scalable potential for civic-tech and public policy applications