qshub / workflow /step2_classify.py
basel00omar's picture
Enforce PDF type domain isolation in classification to prevent structural pages matching architectural categories
b9ca909
Raw History Blame Contribute Delete
3.89 kB
"""
STEP 2 — Classify Pages
AI reads each page title/text and assigns drawing type
User can correct any misclassification
"""
import pandas as pd
import os, io, re, json
from pdf_engine.smart_classifier import classify_all_pages, PAGE_ITEMS_MAP
from pdf_engine.pdf_loader import page_to_pil
from workflow.workflow_state import mark_step_done
# Candidate types per PDF kind — used by the AI-vision refiner
_STRUCT_TYPES = ["foundations", "tie_beam", "column_layout", "upper_columns", "neck_columns",
"columns_1f", "columns_2f", "columns_roof",
"slab_1st", "slab_2nd", "roof_slab"]
_ARCH_TYPES = ["ground_floor_plan", "first_floor_plan", "second_floor_plan",
"roof_floor_plan", "elevations", "setting_out", "schedules"]
def _ai_classify(img_input, pdf_type: str):
"""Classify ONE page by vision when keyword matching is unsure. Returns a
PAGE_ITEMS_MAP key or None."""
from utils.key_manager import get_key_manager
mgr = get_key_manager()
api_key, current_model = mgr.get_key_and_model()
if not api_key or api_key == "NO_API_KEY_FOUND":
return None
choices = _STRUCT_TYPES if pdf_type == "structural" else _ARCH_TYPES
try:
from google import genai
from google.genai import types
import time
import io as _io
if isinstance(img_input, bytes):
img = img_input
elif isinstance(img_input, str) and os.path.exists(img_input):
with open(img_input, "rb") as f:
img = f.read()
elif isinstance(img_input, np.ndarray):
pil_img = page_to_pil(img_input)
buf = _io.BytesIO()
pil_img.save(buf, format="PNG")
img = buf.getvalue()
else:
return None
prompt = (
f"This is a construction drawing. Classify it as one of: {choices}. "
"Reply ONLY with valid JSON: {\"type\": \"<chosen_type>\"}"
)
for attempt in range(2):
api_key, current_model = mgr.get_key_and_model()
if not api_key or api_key == "NO_API_KEY_FOUND":
return None
try:
client = genai.Client(api_key=api_key, http_options={'timeout': 10})
resp = client.models.generate_content(
model=current_model,
contents=[types.Part.from_bytes(data=img, mime_type="image/png"), prompt])
raw = re.sub(r"```json|```", "", resp.text).strip()
result = json.loads(raw).get("type")
return result if result in PAGE_ITEMS_MAP else None
except Exception as e:
last_err = str(e)
if any(code in last_err for code in ("429", "RESOURCE_EXHAUSTED", "quota")):
mgr.mark_rate_limited(api_key, current_model)
continue
time.sleep(1)
return None
except Exception:
return None
def _classify_single(text: str, pdf_type: str) -> dict:
from pdf_engine.smart_classifier import PAGE_ITEMS_MAP
text_lower = text.lower()
scores = {}
for ptype, config in PAGE_ITEMS_MAP.items():
# Enforce PDF domain isolation (structural vs architectural)
if config.get("pdf_type") and config.get("pdf_type") != pdf_type:
continue
score = sum(1 for kw in config.get("drawing_keywords", []) if kw.lower() in text_lower)
scores[ptype] = score
if not scores:
return {"detected_type": "unknown", "confidence": "low", "items": []}
best = max(scores, key=scores.get)
score = scores[best]
return {
"detected_type": best if score > 0 else "unknown",
"confidence": "high" if score >= 1 else "low",
"items": PAGE_ITEMS_MAP.get(best, {}).get("extract_items", []) if score > 0 else [],
}