The latest interface of Gemini Pro supports to send a sequence of interlaced text / files as input (aka parts). The raw JSON API is convenient and straightfoward. One can wrap a utility function in 20 lins of codes.
Utility function
import requests
PROJECT_ID = "test-project-gdg-gemini"
LOCATION = "us-central1"
API_ENDPOINT = f"{LOCATION}-aiplatform.googleapis.com"
MODEL_ID="gemini-1.0-pro-vision"
def request_gemini(prompt, gs_path):
url = f'https://{API_ENDPOINT}/v1/projects/{PROJECT_ID}/locations/{LOCATION}/publishers/google/models/{MODEL_ID}:generateContent'
r = requests.post(
url,
headers={
"Authorization": f'Bearer {access_token}',
"Content-Type": "application/json",
},
json={
"contents": {
"role": "USER",
"parts": [
{"text": prompt},
{"file_data": {
"mime_type": "text/html",
"file_uri": gs_path
}}
]
},
"generation_config": {
"temperature": 0.2,
"top_p": 0.1,
"top_k": 16,
"max_output_tokens": 2048,
"candidate_count": 1,
"stop_sequences": []
},
"safety_settings": {
"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
"threshold": "BLOCK_NONE",
}
})
return r.json()
def chat(prompt, gs_path):
r = request_gemini(prompt, gs_path)
t = r['candidates'][0]['content']['parts'][0]['text']
return t
def transfer_to_gs(url, gs_filename):
r = requests.get(url)
open(gs_filename, 'wb').write(r.content)
!gsutil cp {gs_filename} gs://test-gemini-files/
Caution:
- Change the project to yours.
- The
!gsutil cp is a shortcut in CoLab/ Jupyter env to copy files to Google Cloud Storage. You may need to change to proper Python code in prod environment.
- Assign access token with the auth method suitable in your env. One example in CoLab is as below.
access_token = !gcloud auth print-access-token
access_token = access_token[0]
Usage
url = 'https://hupili.net/article/ultra-notes-on-track/'
transfer_to_gs(url, 'on-track.html')
# Checkout the head of the downloaded file
!head -n30 on-track.html
#
chat('prompt here', 'gs://test-gemini-files/on-track.html')
Below are some sample chat logs:
Notes
- The file needs to be stored on Google Cloud Storage.
- Each chat takes about 10-20 seconds to respond.
