TrustVote AI

Appearance

Core Data Schema ​

Context ​

We need a robust database structure to store legislative documents, their vector embeddings for AI analysis, and the voting records. The schema must be optimized for PostgreSQL with pgvector.

Proposed Schema ​

1. Table: legislative_docs ​

Stores the raw and processed text of laws/proposals.

  • id: UUID (Primary Key)
  • title: String
  • content: Text (The full law text)
  • url: String (Source link)
  • status: Enum (Draft, Active, Revoked)
  • created_at: Timestamp

2. Table: doc_chunks_embeddings ​

This is where the AI/LangChain will live. We split large laws into "chunks" for better RAG performance.

  • id: UUID
  • doc_id: ForeignKey(legislative_docs)
  • chunk_content: Text (A portion of the law)
  • embedding: Vector(1536) <-- Optimized for OpenAI/Open-source models
  • metadata: JSONB (Page number, section title)

3. Table: votes ​

The integrity core.

  • id: UUID
  • doc_id: ForeignKey(legislative_docs)
  • user_id: UUID
  • decision: Enum (Yes, No, Abstain)
  • hash: String (SHA-256 of the vote data for integrity)
  • timestamp: Timestamp (Immutable)

Rationale ​

  • Chunking Strategy: Large laws cannot be sent to an LLM at once. Storing chunks with embeddings allows for precise "Semantic Search".
  • JSONB Metadata: Allows flexibility for different types of legislation without changing the schema.
  • Vector(1536): Standard size for many high-performance embedding models.