kb

name	kb
description	Knowledge base and knowledge graph for NPA property intelligence with temporal tracking. Every ingestion records category, area, source, and auto-calculated expiry. Powered by LightRAG (vector + graph) + kb_metadata table for freshness management.

Knowledge Base Skill

Overview

Persistent knowledge storage with temporal metadata tracking. Every document ingested is tagged with a category, area, source, and auto-calculated expiry date. This ensures NPA-guy can detect stale data (old pricing, outdated rental rates) and re-verify before making recommendations.

Capabilities

1. Insert Document (with temporal metadata)

• Method: insert_document(content, description, category, area, source)
  • content (str): Document text (recommended under 2000 chars)
  • description (str): Brief label
  • category (str): Required — determines expiry (see TTL table below)
  • area (str): Geographic area (e.g. "สุขุมวิท 77", "อ่อนนุช")
  • source (str): Data source (e.g. "DDProperty", "Hipflat", "web_search")
  • Returns: Success message with metadata confirmation

2. Query Knowledge

• Method: query_knowledge(query, mode)
  • Results include [Date: ...] headers — check freshness before relying on data
  • Modes: hybrid (default), local, global, mix, naive

3. Check Freshness

• Method: check_freshness(area, category)
  • Returns: Count of fresh vs stale entries for an area/category
  • Use before making pricing or rental recommendations

4. Get Stale Entries

• Method: get_stale_entries(limit)
  • Returns: List of expired entries that need re-verification
  • Use during heartbeat to identify what needs re-searching

5. Graph Statistics + Health Check

• get_graph_stats() — node/edge counts + metadata summary
• health_check() — LightRAG + metadata table status

Category TTL (Time-to-Live)

Category	TTL	Use For
pricing	90 days	Sale prices, price/sqm benchmarks, appraisal comparisons
rental	90 days	Rental rates, yield data, vacancy info
flood	365 days	Flood reports, risk assessments, drainage info
legal	180 days	Title issues, encumbrances, court findings
area	180 days	Area intelligence, amenities, neighborhood quality
project	365 days	Developer info, building reviews, juristic person
infrastructure	365 days	BTS extensions, new roads, malls, development plans
other	180 days	Default for anything else

Ingestion Format

Always include category, area, and source:

insert_document(
    content="On Nut area condos rent 12,000-15,000 THB/month for 30-35 sqm units.",
    description="On Nut condo rental rates April 2026",
    category="rental",
    area="อ่อนนุช",
    source="DDProperty"
)

Freshness Checker CLI

# Full report: freshness by category, area, coverage gaps
python scripts/kb_freshness.py

# Only stale and expiring-soon documents
python scripts/kb_freshness.py --stale

# Detailed report for a specific area
python scripts/kb_freshness.py --area "บางเขน"

# Prioritized re-research queue (areas/categories to refresh)
python scripts/kb_freshness.py --queue

Legacy Staleness Checker

python scripts/check_stale.py                     # List stale entries
python scripts/check_stale.py --summary            # Quick summary by category
python scripts/check_stale.py --mark               # Mark expired as stale
python scripts/check_stale.py --category pricing   # Check specific category

Query Modes

Mode	Best For
hybrid	Most property analysis queries (default)
local	Specific properties, projects, locations
global	Area trends, market patterns
mix	Finding connections between entities
naive	Quick keyword-like search

What to Ingest vs What NOT to Ingest

INGEST (evolving knowledge with category tag):

• Property analysis results → category: area
• Web search: sale prices → category: pricing
• Web search: rental rates → category: rental
• Web search: flood reports → category: flood
• Legal findings → category: legal
• Project/developer reviews → category: project
• Infrastructure plans → category: infrastructure

DO NOT INGEST (static reference data already in scripts):

• Zoning rules / FAR / OSR → zoning-check skill
• Tax rates / transfer fees → property-calc skill
• Flood risk zones → flood-check skill
• BTS/MRT station data → location-intel skill

Notes

• Database: npa_kb (separate from Ada's ada_kb)
• Metadata table: kb_metadata in same database
• LLM: Gemini gemini-3.1-flash-lite-preview for entity extraction
• Embeddings: Gemini gemini-embedding-001 (3072 dimensions)
• Temporal headers are prepended to content for LightRAG entity extraction