48cef99257
Apps: - dwg-rooms: extract room numbers from DWG/DXF - dwg-counting: count symbols in PDF drawings (OpenCV template matching) - contract-check: review PDF contracts against a checklist (Claude vision + Tesseract OCR fallback) - email-drafter: bullet notes → polished Czech/English business emails - invoice-extractor: PDF/image invoice → structured data → Excel - translator: Czech-first translator across 19 languages with tone control - vv-check: find inconsistent unit prices across VV sheets in one workbook - vv-compare: diff original vs new VV files (changes / added / removed) - feature-request: portal users submit ideas + sample files Infrastructure: - LiteLLM gateway with per-app virtual keys + budgets - Langfuse observability - Geist font, shared theme, cross-subdomain back link + theme sync via cookie/URL - Caddy reverse proxy on *.klas.chat
81 lines
2.7 KiB
Python
81 lines
2.7 KiB
Python
"""LLM fallback: classify unmatched DXF text entities as rooms via LiteLLM."""
|
|
import json
|
|
import logging
|
|
import os
|
|
|
|
from openai import AsyncOpenAI
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
_client: AsyncOpenAI | None = None
|
|
|
|
|
|
def _get_client() -> AsyncOpenAI:
|
|
global _client
|
|
if _client is None:
|
|
_client = AsyncOpenAI(
|
|
base_url=os.getenv("LITELLM_BASE_URL", "http://host.docker.internal:4000"),
|
|
api_key=os.getenv("LITELLM_API_KEY", "sk-dummy"),
|
|
)
|
|
return _client
|
|
|
|
|
|
SYSTEM = """You are a specialist extracting room data from Czech architectural DXF floor plans.
|
|
You receive text entities (text, x, y) that were not matched by rule-based parsing.
|
|
|
|
Identify pairs of room number + Czech room name/description.
|
|
Czech room numbers: 4-6 digit codes, sometimes prefixed with "č.m.".
|
|
Czech room names: e.g. "Chodba", "Serverovna", "Sklep", "WC", "Kancelář", etc.
|
|
|
|
Return ONLY a JSON array of objects:
|
|
{"room": "XXXXX", "description": "Czech name", "confidence": 0.0-1.0}
|
|
|
|
Skip: measurements (m2, m²), material names (beton, dlažba), dimensions, unrelated text.
|
|
Only include entries with confidence > 0.5."""
|
|
|
|
|
|
async def enhance_with_llm(unmatched: list[dict]) -> list[dict]:
|
|
api_key = os.getenv("LITELLM_API_KEY", "")
|
|
if not api_key or api_key == "sk-dummy":
|
|
logger.info("LITELLM_API_KEY not set — skipping LLM enhancement")
|
|
return []
|
|
|
|
sample = unmatched[:200]
|
|
text_block = "\n".join(
|
|
f'- "{t["text"]}" x={t["x"]:.0f} y={t["y"]:.0f}' for t in sample
|
|
)
|
|
model = os.getenv("LLM_MODEL", "gpt-4o-mini")
|
|
|
|
try:
|
|
resp = await _get_client().chat.completions.create(
|
|
model=model,
|
|
messages=[
|
|
{"role": "system", "content": SYSTEM},
|
|
{"role": "user", "content": f"Text entities:\n{text_block}"},
|
|
],
|
|
temperature=0.1,
|
|
max_tokens=2000,
|
|
)
|
|
raw = resp.choices[0].message.content or "[]"
|
|
# Strip markdown code fences if present
|
|
raw = raw.strip().removeprefix("```json").removeprefix("```").removesuffix("```").strip()
|
|
data = json.loads(raw)
|
|
if isinstance(data, dict):
|
|
data = data.get("rooms", data.get("result", []))
|
|
|
|
return [
|
|
{
|
|
"room": str(r["room"]),
|
|
"description": r.get("description", ""),
|
|
"x": 0.0,
|
|
"y": 0.0,
|
|
"source": "llm",
|
|
"confidence": float(r.get("confidence", 0.7)),
|
|
}
|
|
for r in data
|
|
if isinstance(r, dict) and r.get("room")
|
|
]
|
|
except Exception as exc:
|
|
logger.error("LLM enhancement failed: %s", exc)
|
|
return []
|