使用 BigQuery 和 Gemini 模型创建营销活动

picture_description = "Mocha Pancakes with Vanilla Bean ice cream"

llm_human_image_prompt = f"Create an image that contains a {picture_description} that I can use for a marketing campaign."

human_prompt_filename = ImageGen(llm_human_image_prompt)

img = Image.open(human_prompt_filename)

llm_human_image_json = {

   "gcs_storage_bucket" : gcs_storage_bucket,

   "gcs_storage_path" : gcs_storage_path,

   "llm_image_filename" : human_prompt_filename

}

img.thumbnail([500,500])

IPython.display.display(img)

example_return_json = '[ { "prompt" : "response 1" }, { "prompt" : "response 2" }, { "prompt" : "A prompt for good LLMs" }, { "prompt" : "Generate an image that is awesome" }]'

llm_generated_image_prompt=f"""For the below prompt, rewrite it in 10 different ways so I get the most creative images to be generated.

Try creative thinking, generate innovative and out-of-the-box ideas to solve the problem.

Explore unconventional solutions, thinking beyond traditional boundaries, and encouraging imagination and originality.

Embrace unconventional ideas and mutate the prompt in a way that surprises and inspires unique variations.

Think outside the box and develop a mutator prompt that encourages unconventional approaches and fresh perspectives.

Return the results in JSON with no special characters or formatting.

Limit each json result to 256 characters.

Example Return Data:

{example_return_json}

Prompt: "Create a picture of a {picture_description} to be used for a marketing campaign."

"""

llm_success = False

temperature=.8

while llm_success == False:

 try:

   llm_response = GeminiProLLM(llm_generated_image_prompt, temperature=temperature, topP=.8, topK = 40)

   llm_success = True

 except:

   # Reduce the temperature for more accurate generation

   temperature = temperature - .05

   print("Regenerating...")

print(llm_response)

llm_json = json.loads(llm_response)

# Add an image to the generation that will not contain any food items.  We will test this later.

llm_json.append({'prompt' : 'Draw a coffee truck with disco lights.'})

image_files = []

for item in llm_json:

 print(item["prompt"])

 try:

   image_file = ImageGen(item["prompt"])

   image_files.append ({

       "llm_prompt" : item["prompt"],

       "gcs_storage_bucket" : gcs_storage_bucket,

       "gcs_storage_path" : gcs_storage_path,

       "llm_image_filename" : image_file,

       "llm_validated_by_llm" : False

   })

 except:

   print("Image failed to generate.")

example_return_json = '{ "response" : true, "explanation" : "Reason why..." }'

llm_validated_image_json = []

number_of_valid_images = 0

llm_validated_image_prompt = f"""Is attached image a picture of "{picture_description}"?

Respond with a boolean of true or false and place in the "response" field.

Explain your reasoning for each image and place in the "explanation" field.

Return the results in JSON with no special characters or formatting.

Place the results in the following JSON structure:

{example_return_json}

"""

for item in image_files:

 print(f"LLM Prompt : {item['llm_prompt']}")

 print(f"LLM Filename : {item['llm_image_filename']}")

 imageBase64 = convert_png_to_base64(item['llm_image_filename'])

 llm_success = False

 temperature = .4

 while llm_success == False:

   try:

     llm_response = GeminiProVisionLLM(llm_validated_image_prompt, imageBase64, temperature=temperature)

     print(f"llm_response : {llm_response}")

     llm_json = json.loads(llm_response)

     llm_success = True

   except:

     # Reduce the temperature for more accurate generation

     temperature = temperature - .05

     print("Regenerating...")

 # Mark this item as useable

 if llm_json["response"] == True:

   item["llm_validated_by_llm"] = True

   number_of_valid_images = number_of_valid_images + 1

 item["llm_validated_by_llm_explanation"] = llm_json["explanation"]

 llm_validated_image_json.append( {

     "llm_prompt" : item["llm_prompt"],

     "gcs_storage_bucket" : item["gcs_storage_bucket"],

     "gcs_storage_path" : item["gcs_storage_path"],

     "llm_image_filename" : item["llm_image_filename"],

     "llm_validated_by_llm" : item["llm_validated_by_llm"],

     "llm_validated_by_llm_explanation" : item["llm_validated_by_llm_explanation"]

     })

image_prompt = []

for item in image_files:

 if item["llm_validated_by_llm"] == True:

   print(f"Adding image {item['llm_image_filename']} to taste test.")

   imageBase64 = convert_png_to_base64(item['llm_image_filename'])

   new_item = {

       "llm_image_filename" : item['llm_image_filename'],

       "llm_image_base64" : f"{imageBase64}"

   }

   image_prompt.append(new_item)

example_return_json='[ {"image_name" : "name", "rating" : 10, "explanation": ""}]'

llm_taste_test_image_prompt=f"""You are going to be presented with {number_of_valid_images}.

You must critique each image and assign it a score from 1 to 100.

You should compare the images to one another.

You should evaluate each image multiple times.

Score each image based upon the following:

- Appetizing images should get a high rating.

- Realistic images should get a high rating.

- Thought provoking images should get a high rating.

- Plastic looking images should get a low rating.

- Abstract images should get a low rating.

Think the rating through step by step for each image.

Place the result of the scoring process in the "rating" field.

Return the image name and place in the "image_name" field.

Explain your reasoning and place in the "explanation" field in less than 20 words.

Place the results in the following JSON structure:

{example_return_json}

"""

llm_success = False

temperature=1

while llm_success == False:

 try:

   llm_response = GeminiProVisionMultipleFileLLM(llm_taste_test_image_prompt, image_prompt, temperature = temperature)

   llm_success = True

 except:

   # Reduce the temperature for more accurate generation

   temperature = temperature - .05

   print("Regenerating...")

print(f"llm_response : {llm_response}")

example_return_json_1 = '{ "city" : "New York City", "subject" : "email subject" , "body" : "email body" }'

example_return_json_2 = '{ "city" : "London", "subject" : "marketing subject" , "body" : "The body of the email message" }'

imageBase64 = convert_png_to_base64(highest_rated_llm_image_filename)

llm_marketing_campaign = []

# Loop for each city

for city_index in range(0, 4):

 print(f"Processing city: {city_names[city_index]}")

 prompt=f"""You run a fleet of coffee trucks in {city_names[city_index]}.

 We need to craft a compelling email marketing campaign for the below image.

 Embrace unconventional ideas and thinking that surprises and inspires unique variations.

 We want to offer exclusive discounts or early access to drive sales.

 Look at the image and come up with a savory message.

 The image contains "{picture_description}".

 The marketing campaign should be personalized for {city_names[city_index]}.

 Do not mention the weather in the message.

 Write the response using the language "{city_languages[city_index]}".

 The campaign should be less than 1500 characters.

 The email message should be formatted in text, not HTML.

 Sign the email message "Sincerely, Data Beans".

 Mention that people can find the closest coffee truck location by using our mobile app.

 Return the results in JSON with no special characters or formatting.

 Double check for special characters especially for Japanese.

 Place the value of "{city_names[city_index]}" in the "city" field of the JSON response.

 Example Return Data:

 {example_return_json_1}

 {example_return_json_2}

 Image of items we want to sell:

 """

 llm_success = False

 temperature = .8

 while llm_success == False:

   try:

     llm_response = GeminiProVisionLLM(prompt, imageBase64, temperature=.8, topP=.9, topK = 40)

     llm_json = json.loads(llm_response)

     print(llm_response)

     llm_success = True

   except:

     # Reduce the temperature for more accurate generation

     temperature = temperature - .05

     print("Regenerating...")

 llm_marketing_campaign.append(llm_json)

 llm_marketing_prompt.append(prompt)

 llm_marketing_prompt_text.append(llm_response)

imageBase64 = convert_png_to_base64(highest_rated_llm_image_filename)

for item in llm_marketing_campaign:

 print (f'City:    {item["city"]}')

 print (f'Subject: {item["subject"]}')

 prompt=f"""Convert the below text into a well-structured HTML document.

Refrain from using <h1>, <h2>, <h3>, and <h4> tags.

Create inline styles.

Make the styles fun, eye-catching and exciting.

Use "Helvetica Neue" as the font.

All text should be left aligned.

Avoid fonts larger than 16 pixels.

Do not change the language.  Keep the text in the native language.

Include the following image in the html:

- The image is located at: https://REPLACE_ME

- The image url should have a "width" of 500 and "height" of 500.

Double check that you did not use any <h1>, <h2>, <h3>, or <h4> tags.

Text:

{item["body"]}

"""

 llm_success = False

 temperature = .5

 while llm_success == False:

   try:

     llm_response = GeminiProLLM(prompt, temperature=temperature, topP=.5, topK = 20)

     if llm_response.startswith("html"):

       llm_response = llm_response[4:] # incase the response is formatted like markdown

     # Replace the image with an inline image (this avoids a SignedURL or public access to GCS bucket)

     html_text = llm_response.replace("https://REPLACE_ME", f"data:image/png;base64, {imageBase64}")

     item["html"] = html_text

     #print (f'Html:    {item["html"]}')

     filename= str(item["city"]).replace(" ","-") + ".html"

     with open(filename, "w", encoding='utf8') as f:

       f.write(item["html"])

     print ("")

     llm_success = True

     llm_marketing_prompt_html.append(prompt)

     llm_marketing_prompt_html_text.append(html_text)

   except:

     # Reduce the temperature for more accurate generation

     temperature = temperature - .05

     print("Regenerating...")