AgentSkillsCN

rag-architect

负责Property Lens的RAG系统设计，涵盖向量数据库管理、DLD文档分块，以及混合检索的落地实施。

SKILL.md
--- frontmatter
name: rag-architect
description: Design of the RAG system for Property Lens. Handles vector DB management, DLD document chunking, and hybrid search implementation.

RAG Architect Skill

Retrieval Strategy

  1. •Hybrid Search: Keyword search (BM25) for building names + Vector search for semantic queries.
  2. •Metadata Filtering: Hard filter by District and Transaction Date BEFORE performing vector search.
  3. •Re-ranking: Utilize Cross-encoders to rank the top 5 most relevant transactions.

Data Processing

  • •Chunking: 512-token chunks with 10% overlap for legal/DLD texts.
  • •Embedding Model: Utilize text-embedding-3-small or higher density models for financial data.

Output Validation

  • •Source Attribution: Every AI-generated claim must include a reference to the source data point.