"""Spike Test 1 — 공통 옥탑방 base floor plan spec.json LLM 출력 검증. 목적: GPT-5.5 + Gemini-3.1-pro에게 옥탑방 씬 텍스트 + selected shot description 전달 → spec.json (rooms polygon + doors + furniture) 출력 → schema validation + 좌표 합리성 검증 비교 대상: 실험 v3 수동 도면 (FLOORPLAN_V3_PROMPT) — 사람이 시나리오 분석 + 3차례 정제한 결과. LLM이 그 수준의 정확도로 spec.json을 한 번에 만들어내는지가 핵심 검증 항목. 실행: backend cwd에서 .venv/bin/python scripts/spike_floor_plan_svg/test_01_base_floor_plan.py """ from __future__ import annotations import concurrent.futures import json import os import sys import urllib.request from pathlib import Path from typing import Any, Dict, List try: from dotenv import load_dotenv load_dotenv(Path(__file__).parent.parent.parent / ".env") except ImportError: pass PID = "c00bbe19-a9b5-463f-acfc-806f2e820258" EID = "fe165e3a-19c2-4a0f-9acb-e0c9bab0ee5a" ROOT = Path("/Users/manta/Documents/Projects/TheRoad-I1") CKPT = ROOT / "projects" / PID / "checkpoints" / "episodes" / EID OUT = Path(__file__).parent / "results" OUT.mkdir(parents=True, exist_ok=True) ROOFTOP_SCENES = (5, 12, 14, 18, 25, 27) # ───────── spec.json schema ───────── SPEC_SCHEMA: Dict[str, Any] = { "type": "object", "properties": { "schema_version": {"type": "integer"}, "location_id": {"type": "string"}, "location_type": { "type": "string", "description": "Generic (no scenario words). e.g., 'rooftop_apartment_two_room'", }, "canvas": { "type": "object", "description": "Logical coordinate canvas. Top-down view. Origin (0,0) at bottom-left.", "properties": { "width": {"type": "number"}, "height": {"type": "number"}, "unit": {"type": "string", "description": "e.g., 'meters' or 'logical'"}, }, "required": ["width", "height", "unit"], "additionalProperties": False, }, "rooms": { "type": "array", "items": { "type": "object", "properties": { "id": {"type": "string", "description": "snake_case, e.g., 'living'"}, "label_ko": {"type": "string"}, "label_en": {"type": "string"}, "polygon": { "type": "array", "description": "ordered vertices [[x,y],...] forming closed polygon", "items": { "type": "array", "items": {"type": "number"}, "minItems": 2, "maxItems": 2, }, "minItems": 3, }, }, "required": ["id", "label_ko", "label_en", "polygon"], "additionalProperties": False, }, }, "doors": { "type": "array", "items": { "type": "object", "properties": { "id": {"type": "string"}, "from_room": {"type": "string", "description": "room.id or 'exterior'"}, "to_room": {"type": "string", "description": "room.id or 'exterior'"}, "position": { "type": "array", "items": {"type": "number"}, "minItems": 2, "maxItems": 2, }, "kind": { "type": "string", "enum": ["interior", "front_entry", "window"], }, }, "required": ["id", "from_room", "to_room", "position", "kind"], "additionalProperties": False, }, }, "furniture": { "type": "array", "items": { "type": "object", "properties": { "id": {"type": "string"}, "type": { "type": "string", "description": "generic: tv, sofa, table, sink, bed, mirror, toilet, curtain, etc.", }, "room": {"type": "string"}, "wall": { "type": "string", "enum": ["left", "right", "top", "bottom", "interior", ""], }, "position": { "type": "array", "items": {"type": "number"}, "minItems": 2, "maxItems": 2, }, "size": { "type": "array", "items": {"type": "number"}, "minItems": 2, "maxItems": 2, }, "label_en": {"type": "string"}, }, "required": ["id", "type", "room", "wall", "position", "size", "label_en"], "additionalProperties": False, }, }, "exterior_adjacency_zones": { "type": "array", "items": { "type": "object", "properties": { "id": {"type": "string"}, "type": { "type": "string", "description": "e.g., 'rooftop_concrete', 'street', 'balcony'", }, "polygon": { "type": "array", "items": { "type": "array", "items": {"type": "number"}, "minItems": 2, "maxItems": 2, }, "minItems": 3, }, "connection_door": {"type": "string", "description": "door.id"}, "label_en": {"type": "string"}, }, "required": ["id", "type", "polygon", "connection_door", "label_en"], "additionalProperties": False, }, }, "rejected_layouts": { "type": "array", "items": {"type": "string"}, "description": "Layout patterns that contradict the scenario. Avoid generating these.", }, }, "required": [ "schema_version", "location_id", "location_type", "canvas", "rooms", "doors", "furniture", "exterior_adjacency_zones", "rejected_layouts", ], "additionalProperties": False, } # ───────── Prompt ───────── SYSTEM_PROMPT = """You are an architectural layout planner for cinematic scene rendering. Given the screenplay scenes that occur in a single shooting LOCATION, output a structured `floor_plan_spec` JSON describing the location's top-down geometry. CRITICAL RULES: 1. Coordinate system: top-down view. Origin (0,0) at bottom-left. X increases rightward, Y increases upward. Use meters as units. 2. Output ONLY structured JSON conforming to the schema. No prose. 3. Do NOT use any scenario-specific proper nouns (character names, place names, work titles) in id/label_en/type/rejected_layouts. Korean labels (label_ko) MAY use Korean common nouns (e.g., '거실', '안방') but never proper nouns. 4. Every furniture item must lie INSIDE its assigned room polygon. 5. Every door must lie ON a room boundary edge or wall. 6. exterior_adjacency_zones describe what is OUTSIDE the building near a front_entry door (e.g., rooftop concrete floor, street, balcony). 7. rejected_layouts: list 2~5 layout patterns that the scenario explicitly contradicts. Examples: "single-room studio (the scenario shows separate bedrooms)", "balcony connection (the scenario shows rooftop access)". 8. Read every scene/shot carefully. Identify zone markers (e.g., '/거실', '/안방', '/욕실', '/현관'). The presence of those markers means the location has those zones. 9. Identify external connection from front entry: is it a rooftop? hallway? street? — match the scenario. OUTPUT REQUIREMENTS: - 5~7 rooms typically (interior zones) - 1 front_entry door + N interior doors - 6~12 furniture items minimum (the visible/important ones from the scenes/shots) - 1+ exterior_adjacency_zones if front entry leads outside """ def _build_user_prompt() -> str: """옥탑방 씬 텍스트 + selected shot description 빌드.""" scene_save = json.loads((CKPT / "scene_save" / "manifest.json").read_text()) sv = json.loads((CKPT / "shot_validator" / "manifest.json").read_text())["data"] ss = json.loads((CKPT / "shot_selection" / "manifest.json").read_text())["data"] segs = scene_save["data"]["segments"] sel_map = {s["scene_index"]: set(s.get("selected_shot_indices", [])) for s in ss.get("scenes", [])} parts: List[str] = [ "LOCATION ID: L05", "LOCATION DESCRIPTION (from scenario): a Korean rooftop apartment that the family lives in.", "", "── ALL SCENES OCCURRING IN THIS LOCATION ──", "", ] for seg in segs: if seg["scene_index"] in ROOFTOP_SCENES: parts.append(f"### Scene {seg['scene_index']} — {seg.get('heading', '')}") parts.append(seg.get("text", "")) parts.append("") parts.append("── SELECTED SHOTS IN THIS LOCATION (visible elements) ──") parts.append("") for s in sv.get("scenes", []): si = s["scene_index"] if si not in ROOFTOP_SCENES: continue sel = sel_map.get(si, set()) parts.append(f"### Scene {si} selected shots:") for sh in s.get("shots", []): if sh["shot_index"] in sel: parts.append(f"- Shot {sh['shot_index']}: {sh.get('description','')}") parts.append("") parts.append("") parts.append("Now output the floor_plan_spec JSON for this location.") return "\n".join(parts) # ───────── LLM clients ───────── def call_gpt(user_prompt: str) -> Dict[str, Any]: from openai import OpenAI client = OpenAI(api_key=os.environ["OPENAI_API_KEY"]) resp = client.chat.completions.create( model="gpt-5.5", messages=[ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": user_prompt}, ], response_format={ "type": "json_schema", "json_schema": { "name": "floor_plan_spec", "schema": SPEC_SCHEMA, "strict": True, }, }, ) return json.loads(resp.choices[0].message.content) def call_gemini(user_prompt: str) -> Dict[str, Any]: """Gemini REST API 직접 호출 — production 패턴 (gemini_text_client.py)과 동일.""" api_key = os.environ.get("GEMINI_API_KEY") or os.environ.get("GOOGLE_API_KEY") if not api_key: raise RuntimeError("GEMINI_API_KEY (or GOOGLE_API_KEY) not set") model = os.environ.get("GEMINI_TEXT_MODEL", "gemini-3.1-pro-preview") url = ( f"https://generativelanguage.googleapis.com/v1beta/models/{model}" f":generateContent?key={api_key}" ) body: Dict[str, Any] = { "systemInstruction": {"parts": [{"text": SYSTEM_PROMPT}]}, "contents": [{"role": "user", "parts": [{"text": user_prompt}]}], "generationConfig": { "temperature": 0.2, "responseMimeType": "application/json", "responseJsonSchema": SPEC_SCHEMA, "maxOutputTokens": 16384, }, } req = urllib.request.Request( url, data=json.dumps(body).encode("utf-8"), headers={"Content-Type": "application/json"}, method="POST", ) with urllib.request.urlopen(req, timeout=600) as resp: payload = json.loads(resp.read().decode("utf-8")) text = payload["candidates"][0]["content"]["parts"][0]["text"] return json.loads(text) # ───────── Validation ───────── def _point_in_polygon(point, polygon) -> bool: """Ray casting algorithm.""" x, y = point inside = False n = len(polygon) j = n - 1 for i in range(n): xi, yi = polygon[i] xj, yj = polygon[j] if ((yi > y) != (yj > y)) and (x < (xj - xi) * (y - yi) / (yj - yi + 1e-12) + xi): inside = not inside j = i return inside def validate_spec(spec: Dict[str, Any]) -> List[str]: """좌표 합리성 검증. 위반 시 경고 list 반환 (빈 list = 통과).""" warnings: List[str] = [] rooms = {r["id"]: r for r in spec.get("rooms", [])} # 1) 가구가 room 안에 있는지 for f in spec.get("furniture", []): room_id = f.get("room") if room_id not in rooms: warnings.append(f"furniture '{f['id']}' room='{room_id}' not in rooms") continue polygon = rooms[room_id]["polygon"] if not _point_in_polygon(f["position"], polygon): warnings.append( f"furniture '{f['id']}' at {f['position']} OUTSIDE room '{room_id}'" ) # 2) 문이 room polygon edge 위 또는 근처(<0.5m)에 있는지 for d in spec.get("doors", []): for room_label in (d.get("from_room"), d.get("to_room")): if room_label == "exterior": continue if room_label not in rooms: warnings.append(f"door '{d['id']}' room='{room_label}' not in rooms") # 3) room polygon 자기 교차 검사 (간단히 vertex 수 ≥ 3) for r in spec.get("rooms", []): if len(r["polygon"]) < 3: warnings.append(f"room '{r['id']}' polygon has <3 vertices") # 4) front_entry door 1개 이상 n_front = sum(1 for d in spec.get("doors", []) if d.get("kind") == "front_entry") if n_front == 0: warnings.append("no front_entry door") elif n_front > 1: warnings.append(f"multiple ({n_front}) front_entry doors") return warnings # ───────── Main ───────── def main() -> int: user_prompt = _build_user_prompt() (OUT / "user_prompt.txt").write_text(user_prompt, encoding="utf-8") print(f"[INFO] user_prompt 길이: {len(user_prompt)}자 → {OUT/'user_prompt.txt'}") print("[INFO] LLM 병렬 호출 (GPT-5.5 + Gemini-3.1-pro)...") with concurrent.futures.ThreadPoolExecutor(max_workers=2) as ex: f_gpt = ex.submit(call_gpt, user_prompt) f_gem = ex.submit(call_gemini, user_prompt) results: Dict[str, Dict[str, Any] | str] = {} for name, fut in (("gpt", f_gpt), ("gemini", f_gem)): try: results[name] = fut.result(timeout=600) except Exception as exc: results[name] = f"ERROR: {type(exc).__name__}: {exc}" for name, spec in results.items(): out_path = OUT / f"spec_{name}.json" if isinstance(spec, dict): out_path.write_text(json.dumps(spec, ensure_ascii=False, indent=2), encoding="utf-8") warns = validate_spec(spec) n_rooms = len(spec.get("rooms", [])) n_doors = len(spec.get("doors", [])) n_furn = len(spec.get("furniture", [])) n_ext = len(spec.get("exterior_adjacency_zones", [])) n_rej = len(spec.get("rejected_layouts", [])) print(f"\n[{name.upper()}] OK → {out_path}") print(f" rooms={n_rooms} doors={n_doors} furniture={n_furn} exterior={n_ext} rejected={n_rej}") print(f" validation warnings: {len(warns)}") for w in warns: print(f" - {w}") else: print(f"\n[{name.upper()}] {spec}") (OUT / f"spec_{name}_error.txt").write_text(str(spec), encoding="utf-8") return 0 if __name__ == "__main__": sys.exit(main())