python-gemini-api 2.4.11

The python package that returns Response of Google Gemini through API.

Gemini API

https://github.com/dsdanielpark/Gemini-API/assets/81407603/e0c11d4f-3fe1-4cbb-ba79-d9f89b637324

An *unofficial Python wrapper, python-gemini-api, is available for users facing frequent authentication issues or unable to use Google Authentication. This wrapper uses cookie values to interact with Google Gemini through reverse-engineering. The project involved a collaboration with Antonio Cheong.

On the official side, Google provides partially free, clean official Gemini APIs and SDKs, which can be accessed and utilized neatly via Python packages, google-generativeai.

[!TIP] | 2024-03-26 | [See Code Examples]

Check out temporarily free Open-source LLM APIs with Open Router. (Free limit: 10 requests/minute)

Contents

• Gemini API
  • What is Gemini? 🔐
  • Installation ✅
  • Authentication ✅
  • Quick Start ✅
  • Usage
    • # 01. Initialization ✅
    • # 02. Generate content
    • # 03. Send request
    • # 04. Text generation
    • # 05. Image generation
    • # 06. Retrieving Images from Gemini Responses
    • # 07. Generate content from images
    • # 08. Generate content using Google Services
    • # 09. Fix context setting RCID
    • # 10. Changing the Selected Response from 0 to n
    • # 11. Generate custom content
  • Further
  • Google Open-source LLMs
    • Open-source LLM, Gemma 🤝
    • Open-source LLM, Code Gemma 🤝
  • Utilize free open-source LLM API through Open Router ✅

What is Gemini?

| Paper | Official Website | Official API | API Documents |

Gemini is a family of generative AI models developed by Google DeepMind that is designed for multimodal use cases. The Gemini API gives you access to the Gemini Pro and Gemini Pro Vision models. In February 2024, Google's Bard service was changed to Gemini.

Overview of Google LLMs

Model	Type	Access	Details
Gemini	Proprietary	API [13]	A proprietary multimodal AI by Google DeepMind, including advanced models such as Gemini Pro and Gemini Pro Vision. Access is restricted to API usage; additional insights may be obtained through the paper and website. [1][2]
Gemma	Open Source	Downloadable Free API	An open-source text-to-text language model suitable for tasks like QA and summarization. Weights are downloadable for on-premises use, and detailed documentation is provided via the paper and website. [3][4]
Code Gemma	Open Source	Downloadable	Designed specifically for programming tasks, this open-source model offers downloadable weights to assist developers with code generation and similar activities. Refer to the associated paper, blog post, and Hugging Face collection for more information. [5][6][7]

What is Python-Gemini-API?

This is a Python wrapper derived from the Bard API project, designed to retrieve responses from Gemini Web in REST format. Synchronous clients are preferred over asynchronous ones for Gemini because of rate limiting and blocking concerns.

Installation 📦

pip install python-gemini-api

pip install git+https://github.com/dsdanielpark/Gemini-API.git

For the updated version, use as follows:

pip install -q -U python-gemini-api

Authentication

1. Visit https://gemini.google.com/
   With browser open, try auto-collecting cookies first.

   from gemini import Gemini
   
   client = Gemini(auto_cookies=True)
   
   # Testing needed as cookies vary by region.
   # client = Gemini(auto_cookies=True, target_cookies=["__Secure-1PSID", "__Secure-1PSIDTS"])
   # client = Gemini(auto_cookies=True, target_cookies="all") # You can pass whole cookies
   
   response = client.generate_content("Hello, Gemini. What's the weather like in Seoul today?")
   print(response.payload)
   

2. (Manually) F12 for browser console → Session: Application → Cookies → Copy the value of some working cookie sets. If it doesn't work, go to step 3.

   Some working cookie sets

   Cookies may vary by account or region.

   First try __Secure-1PSIDCC alone. If it doesn't work, use __Secure-1PSID and __Secure-1PSIDTS. Still no success? Try these four cookies: __Secure-1PSIDCC, __Secure-1PSID, __Secure-1PSIDTS, NID. If none work, proceed to step 3 and consider sending the entire cookie file.

3. (Recommended) Export Gemini site cookies via a browser extension. For instance, use Chrome extension ExportThisCookies, open, and copy the txt file contents.

Further: For manual collection or Required for a few users upon error

4. For manual cookie collection, refer to this image. Press F12 → Network → Send any prompt to Gemini webui → Click the post address starting with "https://gemini.google.com/_/BardChatUi/data/assistant.lamda.BardFrontendService/StreamGenerate" → Headers → Request Headers → Cookie → Copy and Reformat as JSON manually.
5. (Required for a few users upon error) If errors persist after manually collecting cookies, refresh the Gemini website and collect cookies again. If errors continue, some users may need to manually set the nonce value. To do this: Press F12 → Network → Send any prompt to Gemini webui → Click the post address starting with "https://gemini.google.com/_/BardChatUi/data/assistant.lamda.BardFrontendService/StreamGenerate" → Payload → Form Data → Copy the "at" key value. See this image for reference.

[!IMPORTANT] Experiment with different Google accounts and browser settings to find a working cookie. Success may vary by IP and account status. Once connected, a cookie typically remains effective over a month. Keep testing until successful.

Quick Start

Generate content: returns parsed response.

from gemini import Gemini

cookies = {"<key>" : "<value>"} # Cookies may vary by account or region. Consider sending the entire cookie file.
client = Gemini(cookies=cookies) # You can use various args

response = client.generate_content("Hello, Gemini. What's the weather like in Seoul today?")
response.payload

Generate content from image: you can use image as input.

from gemini import Gemini

cookies = {"<key>" : "<value>"}
client = Gemini(cookies=cookies) # You can use various args

response = client.generate_content("What does the text in this image say?", image='folder/image.jpg')
response.payload

[!NOTE] If the generate_content method returns an empty payload, try executing it again without reinitializing the Gemini object.

Usage

Setting language and Gemini version using environment variables:

Setting Gemini response language (Optional): Check supported languages here. Default is English.

import os

os.environ["GEMINI_LANGUAGE"] = "KR"  # Setting Gemini response language (Optional)
os.environ["GEMINI_ULTRA"] = "1"      # Switch to Gemini-advanced response (Experimental, Optional)
# In some accounts, access to Gemini Ultra may not be available. If that's the case, please revert it back to "0".

# 01. Initialization

Please explicitly declare cookies in dict format. You can also enter the path to the file containing the cookie with cookie_fp(*.json, *.txt supported). Check sample cookie files in assets folder.

from gemini import Gemini

cookies = {
    "__Secure-1PSIDCC" : "value",
    "__Secure-1PSID" : "value",
    "__Secure-1PSIDTS" : "value",
    "NID" : "value",
    # Cookies may vary by account or region. Consider sending the entire cookie file.
  }

client = Gemini(cookies=cookies)
# client = Gemini(cookie_fp="folder/cookie_file.json") # (*.json, *.txt) are supported.
# client = Gemini(auto_cookies=True) # Or use auto_cookies paprameter

Auto Cookie Update

For auto_cookie to be set to True, and adjust target_cookies. Gemini WebUI must be active in the browser. The browser_cookie3 enables automatic cookie collection, though updates may not be complete yet.

# 02. Generate content

Returns Gemini's response, but the first one might be empty.

from gemini import Gemini

cookies = {}
client = Gemini(cookies=cookies)

prompt = "Tell me about Large Language Model."
response = client.generate_content(prompt)
print(response.payload)

[!IMPORTANT] DO NOT send same prompt repeatly. If the session connects successfully and generate_content runs well, CLOSE Gemini website. If Gemini web stays open in the browser, cookies may expire faster.

The output of the generate_content function is GeminiModelOutput, with the following structure:

• rcid: returns the response candidate id of the chosen candidate.
• text: returns the text of the chosen candidate.
• code: returns the codes of the chosen candidate.
• web_images: returns a list of web images from the chosen candidate.
• generated_images: returns a list of generated images from the chosen candidate.
• payload: returns the response dictionary, if available. https://github.com/dsdanielpark/Gemini-API/blob/fdf064c57bc1fb47fbbb4b93067618a200e77f62/gemini/src/model/output.py#L16

# 03. Send request

Send request: returns the request's payload and status_code, making debugging easier.

from gemini import Gemini

cookies = {} 
client = Gemini(cookies=cookies) 

response_text, response_status = client.send_request("Hello, Gemini. Tell me about Large Language Models.")
print(response_text)

You can track the total number of requests made by accessing the request_count property within the Gemini class.

# 04. Text generation

Returns text generated by Gemini.

from gemini import Gemini

cookies = {}
client = Gemini(cookies=cookies)

prompt = "Hello, Gemini. Tell me about Large Language Models."
response = client.generate_content(prompt)
print(response.text)

# 05. Image generation

Returns images generated by Gemini.

Async downloader

from gemini import Gemini, GeminiImage

cookies = {}
client = Gemini(cookies=cookies)

response = client.generate_content("Hello, Gemini. Tell me about Large Language Models.")
generated_images = response.generated_images # Check generated images [Dict]

await GeminiImage.save(generated_images, "output", cookies)
# image_data_dict = await GeminiImage.fetch_images_dict(generated_images, cookies)
# await GeminiImage.save_images(image_data_dict, "output")

Further

Display images in IPython You can display the image or transmit it to another application in byte format.

import io
from gemini import Gemini, GeminiImage
from IPython.display import display, Image

cookies = {}
client = Gemini(cookies=cookies)
bytes_images_dict = GeminiImage.fetch_images_dict_sync(generated_images, cookies) # Get bytes images dict

for image_name, image_bytes in bytes_images_dict.items():
    print(image_name)
    image = Image(data=image_bytes)
    display(image)

Sync downloader

from gemini import Gemini, GeminiImage

cookies = {}
client = Gemini(cookies=cookies)

response = client.generate_content("Create illustrations of Seoul, South Korea.")
generated_images = response.generated_images # Check generated images [Dict]

GeminiImage.save_sync(generated_images, save_path="output", cookies=cookies)

# You can use byte type image dict for printing images as follow:
# bytes_images_dict = GeminiImage.fetch_images_dict_sync(generated_images, cookies) # Get bytes images dict
# GeminiImage.save_images_sync(bytes_images_dict, path="output") # Save to dir

Async downloader wrapper

import asyncio
from gemini import GeminiImage

async def save_generated_images(generated_images, save_path="output", cookies=cookies):
    await GeminiImage.save(generated_images, save_path=save_path, cookies=cookies)

# Run the async function
if __name__ == "__main__":
    cookies = {}
    client = Gemini(cookies=cookies)

    response = client.generate_content("Create illustrations of Seoul, South Korea.")

    generated_images = response.generated_images  
    asyncio.run(save_generated_images(generated_images, save_path="output", cookies=cookies))

GeminiImage.save method logic

import asyncio
from gemini import Gemini, GeminiImage

async def save_generated_images(generated_images, save_path="output", cookies=cookies):
    image_data_dict = await GeminiImage.fetch_images_dict(generated_images, cookies)  # Get bytes images dict asynchronously
    await GeminiImage.save_images(image_data_dict, save_path=save_path)  

# Run the async function
if __name__ == "__main__":
    cookies = {}
    client = Gemini(cookies=cookies)

    response = client.generate_content("Create illustrations of Seoul, South Korea.")

    generated_images = response.generated_images 
    asyncio.run(save_generated_images(generated_images, save_path="output", cookies=cookies))

[!NOTE] Use GeminiImage for image processing. web_images works without cookies, but for images like generated_image from Gemini, pass cookies. Cookies are needed to download images from Google's storage. Check the response or use existing cookies variable.

# 06. Retrieving Images from Gemini Responses

Returns images in response of Gemini.

Async downloader

from gemini import Gemini, GeminiImage

cookies = {}
client = Gemini(cookies=cookies)

response = client.generate_content("Give me a picture of Stanford.")
response_images = response.web_images # Check generated images

await GeminiImage.save(response_images, "output")
# image_data_dict = await GeminiImage.fetch_images_dict(response_images)
# await GeminiImage.save_images(image_data_dict, "output")

Further

Sync downloader

from gemini import Gemini, GeminiImage

cookies = {}
client = Gemini(cookies=cookies)

response = client.generate_content("Give me a picture of Stanford.")
response_images = response.web_images # Check response images

GeminiImage.save_sync(response_images, save_path="output")

# You can use byte type image dict as follow:
# bytes_images_dict = GeminiImage.fetch_bytes_sync(response_images) # Get bytes images dict
# GeminiImage.save_images_sync(bytes_images_dict, save_path="output") # Save to path

Async downloader wrapper

import asyncio
from gemini import Gemini, GeminiImage
   
async def save_response_web_imagse(response_images, save_path="output"):
    await GeminiImage.save(response_images, save_path=save_path)

if __name__ == "__main__":
    cookies = {}
    client = Gemini(cookies=cookies)
    response = client.generate_content("Give me a picture of Stanford.")
    response_images = response.web_images  
    asyncio.run(save_response_web_imagse(response_images, save_path="output"))

GeminiImage.save method logic

import asyncio
from gemini import Gemini, GeminiImage

async def save_response_web_imagse(response_images, save_path="output"):
    image_data_dict = await GeminiImage.fetch_images_dict(response_images)  # Get bytes images dict asynchronously
    await GeminiImage.save_images(image_data_dict, save_path=save_path)  

# Run the async function
if __name__ == "__main__":
    cookies = {}
    client = Gemini(cookies=cookies)
    response = client.generate_content("Give me a picture of Stanford.")
    response_images = response.web_images 
    asyncio.run(save_response_web_imagse(response_images, save_path="output"))

# 07. Generate content from images

Takes an image as input and returns a response.

image = 'folder/image.jpg'
# image = open('folder/image.jpg', 'rb').read() # (jpg, jpeg, png, webp) are supported.

# Image file path or Byte-formatted image array
response = client.generate_content("What does the text in this image say?", image=image)
print(response)

# 08. Generate content using Google Services

To begin, you must link Google Workspace to activate this extension via the Gemini web extension. Please refer to the official notice and review the privacy policies for more details.

extention flags

@Gmail, @Google Drive, @Google Docs, @Google Maps, @Google Flights, @Google Hotels, @YouTube

response = client.generate_content("@YouTube Search clips related with Google Gemini")
response.response_dict

Extension description

• Google Workspace

  • Services: @Gmail, @Google Drive, @Google Docs
  • Description: Summarize, search, and find desired information quickly in your content for efficient personal task management.
  • Features: Information retrieval, document summarization, information categorization

• Google Maps

  • Service: @Google Maps
  • Description: Execute plans using location-based information. Note: Google Maps features may be limited in some regions.
  • Features: Route guidance, nearby search, navigation

• Google Flights

  • Service: @Google Flights
  • Description: Search real-time flight information to plan tailored travel itineraries.
  • Features: Holiday preparation, price comparison, trip planning

• Google Hotels

  • Service: @Google Hotels
  • Description: Search for hotels considering what matters most to you, like having a conversation with a friend.
  • Features: Packing for travel, sightseeing, special relaxation

• YouTube

  • Service: @YouTube
  • Description: Explore YouTube videos and ask questions about what interests you.
  • Features: Problem-solving, generating ideas, search, exploring topics

# 09. Fix context setting RCID

You can specify a particular response by setting its Response Candidate ID(RCID).

# Generate content for the prompt "Give me some information about the USA."
response1 = client.generate_content("Give me some information about the USA.")
# After reviewing the responses, choose the one you prefer and copy its RCID.
client.rcid = "rc_xxxx"

# Now, generate content for the next prompt "How long does it take from LA to New York?"
response2 = client.generate_content("How long does it take from LA to New York?")

# However, RCID may not persist. If parsing fails, reset `client.rcid` to None.
# client.rcid = None

# 10. Changing the Selected Response from 0 to n

In Gemini, generate_content returns the first response. This may vary depending on length or sorting. Therefore, you can specify the index of the chosen response from 0 to n as follows. However, if there is only one response, revert it back to 0.

from gemini import GeminiModelOutput

GeminiModelOutput.chosen = 1 # default is 0
response_choice_1 = client.generate_content("Give me some information about the USA.")

# If not all Gemini returns are necessarily plural, revert back to 0 in case of errors.
#  GeminiModelOutput.chosen = 0

# 11. Generate custom content

Parse the response text to extract desired values.

Using Gemini.generate_custom_content, specify custom parsing to extract specific values. Utilize ParseMethod1 and ParseMethod2 by default, and you can pass custom parsing methods as arguments if desired. Refer to custom_parser.py.

# You can create a parser method that takes response_text as the input for custom_parser.
response_text, response_status = client.send_request("Give me some information about the USA.")

# Use custom_parser function or class inheriting from BaseParser
response = client.generate_custom_content("Give me some information about the USA.", *custom_parser)

https://github.com/dsdanielpark/Gemini-API/blob/31b842488bbc5429ad9c74b1d8b00e20d94e8cb1/gemini/client.py#L323

Further

Use rotating proxies via Smart Proxy by Crawlbase

If you want to avoid blocked requests and bans, then use Smart Proxy by Crawlbase. It forwards your connection requests to a randomly rotating IP address in a pool of proxies before reaching the target website. The combination of AI and ML make it more effective to avoid CAPTCHAs and blocks. The argument at the Secure Sockets Layer (SSL) level may need to be added to the header. Use it in conjunction with verify=False.

# Get your proxy url at crawlbase https://crawlbase.com/docs/smart-proxy/get/
proxy_url = "http://xxxxx:@smartproxy.crawlbase.com:8012" 
proxies = {"http": proxy_url, "https": proxy_url}

client = Gemini(cookies=cookies, proxies=proxies, timeout=30, verify=False)
client.session.header["crawlbaseAPI-Parameters"] = "country=US"
client.generate_content("Hello, Gemini. Give me a beautiful photo of Seoul's scenery.")

Reusable session object

For standard cases, use Gemini class; for exceptions, use session objects. When creating a new bot Gemini server, adjust Headers.MAIN.

import requests
from gemini import Gemini, Headers

cookies = {} 

session = requests.Session()
session.headers = Headers.MAIN
for key, value in cookies.items():
    session.cookies.update({key: value})

client = Gemini(session=session) # You can use various args
response = client.generate_content("Hello, Gemini. Tell me about Large Language Model.")

More features

Explore additional features in this document.

If you want to develop your own simple code, you can start from this simple code example.

Google Open-source LLMs

If you have sufficient GPU resources, you can download weights directly instead of using the Gemini API to generate content. Consider Gemma and Code Gemma, an open-source models available for on-premises use.

Open-source LLM, Gemma

Gemma models are Google's lightweight, advanced text-to-text, decoder-only language models, derived from Gemini research. Available in English, they offer open weights and variants, ideal for tasks like question answering and summarization. For more infomation, visit Gemma-7b model card.

How to use Gemma

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("google/gemma-7b")
model = AutoModelForCausalLM.from_pretrained("google/gemma-7b")

input_text = "Write me a poem about Machine Learning."
input_ids = tokenizer(input_text, return_tensors="pt")

outputs = model.generate(**input_ids)
print(tokenizer.decode(outputs[0]))

Open-source LLM, Code Gemma

CodeGemma, which is an official release from Google for code LLMs, was released on April 9, 2024. It provides three models specifically designed for generating and interacting with code. You can explore the Code Gemma models and view the model card for more details.

How to use Code Gemma

from transformers import GemmaTokenizer, AutoModelForCausalLM

tokenizer = GemmaTokenizer.from_pretrained("google/codegemma-7b-it")
model = AutoModelForCausalLM.from_pretrained("google/codegemma-7b-it")

input_text = "Write me a Python function to calculate the nth fibonacci number."
input_ids = tokenizer(input_text, return_tensors="pt")

outputs = model.generate(**input_ids)
print(tokenizer.decode(outputs[0]))

Utilize free open-source LLM API through Open Router

OpenRouter offers temporary free inference for select models. Obtain an API key from Open Router API and check free models at Open Router models. Use models with a 0-dollar token cost primarily; other models may incur charges. See more at free open-source LLM API guide.

Sync client is favored over async for Gemini due to rate limiting and blocking issues, but OpenRouter offers reliable open-source LLMs for async implementation.

from gemini import OpenRouter

OPENROUTER_API_KEY = "<your_open_router_api_key>"
gemma_client = OpenRouter(api_key=OPENROUTER_API_KEY, model="google/gemma-7b-it:free")

prompt = "Do you know UCA academy in Korea? https://blog.naver.com/ulsancoding"
response = gemma_client.create_chat_completion(prompt)
print(response)

# payload = gemma_client.generate_content(prompt)
# print(payload.json())

The free model list includes:

• google/gemma-7b-it:free - google/gemma-7b from Google ***
• mistralai/mistral-7b-instruct:free - mistralai/Mistral-7B-Instruct-v0.1 for instruction from Mistral AI ****
• huggingfaceh4/zephyr-7b-beta:free - HuggingFaceH4/zephyr-7b-beta ***
• openchat/openchat-7b:free - openchat/openchat for chat **
• openrouter/cinematika-7b:free - jondurbin/cinematika-7b-v0.1
• undi95/toppy-m-7b:free - Undi95/Toppy-M-7B
• gryphe/mythomist-7b:free - Gryphe/MythoMist-7b
• nousresearch/nous-capybara-7b:free - NousResearch/Nous-Capybara-7B-V1 from Nous Research

Sponsor

Use Crawlbase API for efficient data scraping to train AI models, boasting a 98% success rate and 99.9% uptime. It's quick to start, GDPR/CCPA compliant, supports massive data extraction, and is trusted by 70k+ developers.

FAQ

First review HanaokaYuzu/Gemini-API and the Official Google Gemini API before using this package. You can find most help on the FAQ and Issue pages.

Issues

Sincerely grateful for any reports on new features or bugs. Your valuable feedback on the code is highly appreciated. Frequent errors may occur due to changes in Google's service API interface. Both Issue reports and Pull requests contributing to improvements are always welcome. We strive to maintain an active and courteous open community.

Contributors

dig the well before you are thirsty.

We would like to express our sincere gratitude to all the contributors.

This package aims to re-implement the functionality of the Bard API, which has been archived for the contributions of the beloved open-source community, despite Gemini's official API already being available.

Contributors to the Bard API and Gemini API.

Further development potential

Modifications to the async client using my logic are needed, along with automatic cookie collection via browser_cookie3, and implementation of other Bard API features (such as code extraction, export to Replit, graph drawing, etc.).

Please note that while reviewing automatic cookie collection, it appears that cookies expire immediately upon sending a request for collection. Efforts to make it more user-friendly were unsuccessful. Also, the _sid value seems to work normally even when returned as None.

Lastly, if the CustomParser and ResponseParser algorithms do not function properly, new parsing methods can be updated through conditional statements in the relevant sections.

I do not plan to actively curate this repository. Please review HanaokaYuzu/Gemini-API first.

Thank you, and have a great day.

Contacts

Core maintainers:

• Antonio Cheong / [e-mail]
  
• Daniel Park / [e-mail]

License ©️

MIT license, 2024. We hereby strongly disclaim any explicit or implicit legal liability related to our works. Users are required to use this package responsibly and at their own risk. This project is a personal initiative and is not affiliated with or endorsed by Google. It is recommended to use Google's official API.

References

1. Introducing GEMINI: Multimodal Generative Models
2. Google DeepMind: GEMINI Introduction
3. GEMMA: A Unified Language Model for Text Generation, Understanding, Translation, Coding, and Math
4. AI at Google: GEMS Documentation
5. CodeGMMA: Large Language Models Can Write Realistic Programming Assignments
6. Announcing CodeGen: Building Better Developers' Tools Using LLMs
7. Google: CodeGen Release
8. acheong08/Bard
9. dsdanielpark/Bard-API
10. HanaokaYuzu/Gemini-API
11. GoogleCloudPlatform/generative-ai
12. OpenRouter
13. Google AI Studio

Warning Users assume full legal responsibility for GeminiAPI. Not endorsed by Google. Excessive use may lead to account restrictions. Changes in policies or account status may affect functionality. Utilize issue and discussion pages.

Requirements

Python 3.7 or higher.