Files
plain-ub-overfork/app/plugins/ai/gemini_core.py
2025-05-03 21:17:27 +05:30

281 lines
8.9 KiB
Python

import asyncio
import io
import logging
import shutil
import time
from functools import wraps
from mimetypes import guess_type
from google.genai.client import AsyncClient, Client
from google.genai.types import (
DynamicRetrievalConfig,
File,
GenerateContentConfig,
GoogleSearchRetrieval,
Part,
SafetySetting,
Tool,
)
from ub_core.utils import get_tg_media_details
from app import BOT, LOGGER, CustomDB, Message, extra_config
logging.getLogger("google_genai.models").setLevel(logging.WARNING)
DB_SETTINGS = CustomDB["COMMON_SETTINGS"]
try:
client: Client | None = Client(api_key=extra_config.GEMINI_API_KEY)
async_client: AsyncClient | None = client.aio
except:
client = async_client = None
async def init_task():
model_info = await DB_SETTINGS.find_one({"_id": "gemini_model_info"}) or {}
if model_name := model_info.get("model_name"):
Settings.TEXT_MODEL = model_name
if image_model := model_info.get("image_model_name"):
Settings.IMAGE_MODEL = image_model
def run_basic_check(function):
@wraps(function)
async def wrapper(bot: BOT, message: Message):
if not extra_config.GEMINI_API_KEY:
await message.reply(
"Gemini API KEY not found."
"\nGet it <a href='https://makersuite.google.com/app/apikey'>HERE</a> "
"and set GEMINI_API_KEY var."
)
return
if not (message.input or message.replied):
await message.reply("<code>Ask a Question | Reply to a Message</code>")
return
await function(bot, message)
return wrapper
def get_response_content(
response, quoted: bool = False, add_sources: bool = True
) -> tuple[str, io.BytesIO | None]:
try:
candidate = response.candidates
parts = candidate[0].content.parts
parts[0]
except (AttributeError, IndexError, TypeError):
LOGGER.info(response)
return "`Query failed... Try again`", None
image_data = None
text = ""
sources = ""
for part in parts:
if part.text:
text += part + "\n"
if part.inline_data:
image_data = io.BytesIO(part.inline_data.data)
image_data.name = "photo.jpg"
if add_sources:
try:
hrefs = [
f"[{chunk.web.title}]({chunk.web.uri})"
for chunk in candidate.grounding_metadata.grounding_chunks
]
sources = "\n\nSources: " + " | ".join(hrefs)
except (AttributeError, TypeError):
sources = ""
final_text = (text.strip() + sources).strip()
if final_text and quoted and "```" not in final_text:
final_text = f"**>\n{final_text}<**"
return final_text, image_data
async def save_file(message: Message, check_size: bool = True) -> File | None:
media = get_tg_media_details(message)
if check_size:
assert getattr(media, "file_size", 0) <= 1048576 * 25, "File size exceeds 25mb."
download_dir = f"downloads/{time.time()}/"
try:
downloaded_file: str = await message.download(download_dir)
uploaded_file = await async_client.files.upload(
file=downloaded_file,
config={
"mime_type": getattr(media, "mime_type", None) or guess_type(downloaded_file)[0]
},
)
while uploaded_file.state.name == "PROCESSING":
await asyncio.sleep(5)
uploaded_file = await async_client.files.get(name=uploaded_file.name)
return uploaded_file
finally:
shutil.rmtree(download_dir, ignore_errors=True)
PROMPT_MAP = {
"video": "Summarize video and audio from the file",
"photo": "Summarize the image file",
"voice": (
"\nDo not summarise."
"\nTranscribe the audio file to english alphabets AS IS."
"\nTranslate it only if the audio is not english."
"\nIf the audio is in hindi: Transcribe it to hinglish without translating."
),
}
PROMPT_MAP["audio"] = PROMPT_MAP["voice"]
async def create_prompts(
message: Message, is_chat: bool = False, check_size: bool = True
) -> list[File, str] | list[str]:
default_media_prompt = "Analyse the file and explain."
input_prompt = message.filtered_input or "answer"
# Conversational
if is_chat:
if message.media:
prompt = message.caption or PROMPT_MAP.get(message.media.value) or default_media_prompt
text_part = Part.from_text(text=prompt)
uploaded_file = await save_file(message=message, check_size=check_size)
file_part = Part.from_uri(file_uri=uploaded_file.uri, mime_type=uploaded_file.mime_type)
return [text_part, file_part]
return [Part.from_text(text=message.text)]
# Single Use
if reply := message.replied:
if reply.media:
prompt = (
message.filtered_input or PROMPT_MAP.get(reply.media.value) or default_media_prompt
)
text_part = Part.from_text(text=prompt)
uploaded_file = await save_file(message=reply, check_size=check_size)
file_part = Part.from_uri(file_uri=uploaded_file.uri, mime_type=uploaded_file.mime_type)
return [text_part, file_part]
return [Part.from_text(text=input_prompt), Part.from_text(text=str(reply.text))]
return [Part.from_text(text=input_prompt)]
@BOT.add_cmd(cmd="llms")
async def list_ai_models(bot: BOT, message: Message):
"""
CMD: LIST MODELS
INFO: List and change Gemini Models.
USAGE: .llms
"""
model_list = [
model.name.lstrip("models/")
async for model in await async_client.models.list(config={"query_base": True})
if "generateContent" in model.supported_actions
]
model_str = "\n\n".join(model_list)
update_str = (
f"<b>Current Model</b>: <code>"
f"{Settings.TEXT_MODEL if "-i" not in message.flags else Settings.IMAGE_MODEL}</code>"
f"\n\n<blockquote expandable=True><pre language=text>{model_str}</pre></blockquote>"
"\n\nReply to this message with the <code>model name</code> to change to a different model."
)
model_info_response = await message.reply(update_str)
model_response = await model_info_response.get_response(
timeout=60, reply_to_message_id=model_info_response.id, from_user=message.from_user.id
)
if not model_response:
await model_info_response.delete()
return
if model_response.text not in model_list:
await model_info_response.edit(f"<code>Invalid Model... Try again</code>")
return
if "-i" in message.flags:
data_key = "image_model_name"
Settings.IMAGE_MODEL = model_response.text
else:
data_key = "model_name"
Settings.TEXT_MODEL = model_response.text
await DB_SETTINGS.add_data({"_id": "gemini_model_info", data_key: model_response.text})
resp_str = f"{model_response.text} saved as model."
await model_info_response.edit(resp_str)
await bot.log_text(text=resp_str, type=f"ai_{data_key}")
SAFETY_SETTINGS = [
# SafetySetting(category="HARM_CATEGORY_UNSPECIFIED", threshold="BLOCK_NONE"),
SafetySetting(category="HARM_CATEGORY_HATE_SPEECH", threshold="BLOCK_NONE"),
SafetySetting(category="HARM_CATEGORY_DANGEROUS_CONTENT", threshold="BLOCK_NONE"),
SafetySetting(category="HARM_CATEGORY_HARASSMENT", threshold="BLOCK_NONE"),
SafetySetting(category="HARM_CATEGORY_SEXUALLY_EXPLICIT", threshold="BLOCK_NONE"),
SafetySetting(category="HARM_CATEGORY_CIVIC_INTEGRITY", threshold="BLOCK_NONE"),
]
SEARCH_TOOL = Tool(
google_search=GoogleSearchRetrieval(
dynamic_retrieval_config=DynamicRetrievalConfig(dynamic_threshold=0.3)
)
)
SYSTEM_INSTRUCTION = (
"Answer precisely and in short unless specifically instructed otherwise."
"\nIF asked related to code, do not comment the code and do not explain the code unless instructed."
)
class Settings:
TEXT_MODEL = "gemini-2.0-flash"
TEXT_CONFIG = GenerateContentConfig(
candidate_count=1,
# max_output_tokens=1024,
response_modalities=["Text"],
system_instruction=SYSTEM_INSTRUCTION,
temperature=0.69,
tools=[],
)
IMAGE_MODEL = "gemini-2.0-flash-exp"
IMAGE_CONFIG = GenerateContentConfig(
candidate_count=1,
# max_output_tokens=1024,
response_modalities=["Text", "Image"],
# system_instruction=SYSTEM_INSTRUCTION,
temperature=0.99,
)
@staticmethod
def get_kwargs(use_search: bool = False, image_mode: bool = False) -> dict:
if image_mode:
return {"model": Settings.IMAGE_MODEL, "config": Settings.IMAGE_CONFIG}
tools = Settings.TEXT_CONFIG.tools
if not use_search and SEARCH_TOOL in tools:
tools.remove(SEARCH_TOOL)
if use_search and SEARCH_TOOL not in tools:
tools.append(SEARCH_TOOL)
return {"model": Settings.TEXT_MODEL, "config": Settings.TEXT_CONFIG}