ponix-generator / llm_wrapper.py
cwhuh's picture
add : refinement logic
586b09a
raw
history blame
4.07 kB
import logging
from PIL import Image
from io import BytesIO
import requests, os, json, time
from google import genai
prompt_base_path = "src/llm_wrapper/prompt"
client = genai.Client(api_key=os.getenv("GEMINI_API_KEY"))
def encode_image(image_source):
"""
์ด๋ฏธ์ง€ ๊ฒฝ๋กœ๊ฐ€ URL์ด๋“  ๋กœ์ปฌ ํŒŒ์ผ์ด๋“  Pillow Image ๊ฐ์ฒด์ด๋“  ๋™์ผํ•˜๊ฒŒ ์ฒ˜๋ฆฌํ•˜๋Š” ํ•จ์ˆ˜.
์ด๋ฏธ์ง€๋ฅผ ์—ด์–ด google.genai.types.Part ๊ฐ์ฒด๋กœ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค.
Pillow์—์„œ ์ง€์›๋˜์ง€ ์•Š๋Š” ํฌ๋งท์— ๋Œ€ํ•ด์„œ๋Š” ์˜ˆ์™ธ๋ฅผ ๋ฐœ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.
"""
try:
# ์ด๋ฏธ Pillow ์ด๋ฏธ์ง€ ๊ฐ์ฒด์ธ ๊ฒฝ์šฐ ๊ทธ๋Œ€๋กœ ์‚ฌ์šฉ
if isinstance(image_source, Image.Image):
image = image_source
else:
# URL์—์„œ ์ด๋ฏธ์ง€ ๋‹ค์šด๋กœ๋“œ
if isinstance(image_source, str) and (
image_source.startswith("http://")
or image_source.startswith("https://")
):
response = requests.get(image_source)
image = Image.open(BytesIO(response.content))
# ๋กœ์ปฌ ํŒŒ์ผ์—์„œ ์ด๋ฏธ์ง€ ์—ด๊ธฐ
else:
image = Image.open(image_source)
# ์ด๋ฏธ์ง€ ํฌ๋งท์ด None์ธ ๊ฒฝ์šฐ (๋ฉ”๋ชจ๋ฆฌ์—์„œ ์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€ ๋“ฑ)
if image.format is None:
image_format = "JPEG"
else:
image_format = image.format
# ์ด๋ฏธ์ง€ ํฌ๋งท์ด ์ง€์›๋˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ ์˜ˆ์™ธ ๋ฐœ์ƒ
if image_format not in Image.registered_extensions().values():
raise ValueError(f"Unsupported image format: {image_format}.")
buffered = BytesIO()
# PIL์—์„œ ์ง€์›๋˜์ง€ ์•Š๋Š” ํฌ๋งท์ด๋‚˜ ๋‹ค์–‘ํ•œ ์ฑ„๋„์„ RGB๋กœ ๋ณ€ํ™˜ ํ›„ ์ €์žฅ
if image.mode in ("RGBA", "P", "CMYK"): # RGBA, ํŒ”๋ ˆํŠธ, CMYK ๋“ฑ์€ RGB๋กœ ๋ณ€ํ™˜
image = image.convert("RGB")
image.save(buffered, format="JPEG")
return genai.types.Part.from_bytes(data=buffered.getvalue(), mime_type="image/jpeg")
except requests.exceptions.RequestException as e:
raise ValueError(f"Failed to download the image from URL: {e}")
except IOError as e:
raise ValueError(f"Failed to process the image file: {e}")
except ValueError as e:
raise ValueError(e)
def run_gemini(
target_prompt: str,
prompt_in_path: str,
output_structure,
img_in_data: str = None,
model: str = "gemini-2.0-flash",
) -> str:
"""
GEMINI API๋ฅผ ๋™๊ธฐ ๋ฐฉ์‹์œผ๋กœ ํ˜ธ์ถœํ•˜์—ฌ ๋ฌธ์ž์—ด ์‘๋‹ต์„ ๋ฐ›์Šต๋‹ˆ๋‹ค.
retry ๋…ผ๋ฆฌ๋Š” ์ œ๊ฑฐ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
"""
with open(os.path.join(prompt_base_path, prompt_in_path), "r", encoding="utf-8") as file:
prompt_dict = json.load(file)
system_prompt = prompt_dict["system_prompt"]
user_prompt_head = prompt_dict["user_prompt"]["head"]
user_prompt_tail = prompt_dict["user_prompt"]["tail"]
user_prompt_text = "\n".join([user_prompt_head, target_prompt, user_prompt_tail])
input_content = [user_prompt_text]
if img_in_data is not None:
encoded_image = encode_image(img_in_data)
input_content.append(encoded_image)
logging.info("Requested API for chat completion response (sync call)...")
start_time = time.time()
# ๋™๊ธฐ ๋ฐฉ์‹: client.models.generate_content(...)
chat_completion = client.models.generate_content(
model=model,
contents=input_content,
config={
"system_instruction": system_prompt,
"response_mime_type": "application/json",
"response_schema": output_structure,
}
)
chat_output = chat_completion.parsed
input_token = chat_completion.usage_metadata.prompt_token_count
output_token = chat_completion.usage_metadata.candidates_token_count
pricing = input_token / 1000000 * 0.1 * 1500 + output_token / 1000000 * 0.7 * 1500
logging.info(
f"[GEMINI] Request completed (sync). Time taken: {time.time()-start_time:.2f}s / Pricing(KRW): {pricing:.2f}"
)
return chat_output, chat_completion