Ranking¶

Ranking has two layers.

src/casita/rank.py is the deterministic sorter. It handles explicit pipeline state, votes, filtered listings, and heuristic score. Human engagement beats a fresh LLM rank because an active conversation is real work.

src/casita/llm.py is the preference ranker. rank_listings builds a compact brief for each listing, adds route summaries, attaches current feedback, and asks Gemini to return every listing with:

• a rank
• a one-sentence reason
• a severity: ok, concerns, or filtered

The ranking policy keeps the personal assumptions: large dogs, SF walkability, Marin drive context, trail or beach access, and practical livability.

Ways This Could Go Further¶

Ranking is deliberately still prompt-centric and Vertex-only. A future version could make policy changes easier to evaluate, compare deterministic and LLM rank movement, or support another model backend.