astraea/README.md

Astraea ⚖️

A constellation of Washington State family-law specialist agents, backed by a local Ollama semantic-RAG pipeline. Eight specialists, each grounded in its own cited corpus of Washington law (RCW), answer questions with inline citations — never from thin air.

For educational purposes only — not legal advice.

The eight specialists

Agent	Domain	Primary source
🧭 Navigator	Court process, residency, intake, routing	RCW 26.09 (overview)
⚖️ Divorce & Dissolution	No-fault dissolution, legal separation, the 90-day wait	RCW 26.09
🧒 Child Custody & Parenting	Parenting plans, best-interest factors, relocation	RCW 26.09 / 26.10
💵 Child Support	Income-shares model, the economic table	RCW 26.19
🕊️ Spousal Maintenance	Maintenance factors & duration	RCW 26.09.090
🏠 Property & Debt	Community vs. separate property, division	RCW 26.16 / 26.09.080
🛡️ Protection Orders	DV, anti-harassment, ERPO	RCW 7.105
🤝 Mediation & ADR	Mandatory mediation, settlement, collaborative law	RCW 26.09.015

Architecture

Flask (gunicorn + nginx)          Ollama host (nightmare, RTX 4080 SUPER)
  app.py ── /api/chat ──────────► granite4.2 (RAG answer, think:false, num_ctx<=16K)
  rag.py  ── embed + retrieve ───► nomic-embed-text-v2-moe (semantic embeddings)
  agents.py (8 specialist defs)
  books/*.md  (8 cited WA-law corpora, chunked -> embedded -> index.json)

The RAG engine chunks each book on heading + sentence boundaries, embeds every chunk with nomic-embed-text-v2-moe, and retrieves the top-5 by cosine similarity. The answer model (granite4.2) is an instruction-follower that answers strictly from the retrieved context and says so when the answer isn't in the corpus.

Deploy (Proxmox LXC)

# On CT 150 "astraea" (10.30.20.160)
/opt/astraea/venv/bin/gunicorn --workers 2 --threads 4 --bind 0.0.0.0:5000 --timeout 240 app:app
# systemd: astraea.service  ·  nginx: /etc/nginx/sites-enabled/astraea (80 -> 5000)

Redeploy: tar czf /tmp/a.tar.gz app.py rag.py agents.py templates/ && scp → proxmox → pct push 150 → extract → systemctl restart astraea.

Reindex (after editing a book): curl -X POST http://<ct>:5000/api/reindex (or restart the service). The vector index is cached at /opt/astraea/index.json and rebuilt only when a book changes.

API

• GET /api/agents — list specialists
• POST /api/chat — {agent_id, message, history?} → {answer, citations, grounded, latency_ms}
• GET /api/health — agent count + index status
• POST /api/reindex — rebuild the vector index

Public

https://astraea.thetempleofdoom.com (Cloudflare fleet tunnel 1aeb1ac0 → 10.30.20.160:80)