Custom AI Model with product documents

[anbu501]

We intend to create an AI model that possesses contextual knowledge about any given product. The model will be trained using various documents associated with the product like requirement documents, Functional Specifications, Technical Design, etc. The trained AI model should maintain the context of the product and provide solutions to product-specific queries or problems. Currently, we are not able to train AI models with huge product docs without losing the context. The LLMs which we used are splitting the docs into multiple portions and the responses are losing context due to that. Also, we are facing token and character limitations when we wanted to use Vertex AI.

Is this possible with vertex AI? If possible how to implement this?

[kvandres]

Good day @anbu501,

Welcome to Google Cloud Community!

Vertex AI will provide the tools and services that can assist you achieve this task but it is important to keep in mind that its success will depend on the structure of your data, complexity of your product and quality of your data. Here is the recommended Machine Learning workflow in GCP that you can implement for your custom trained models based on your data. 

1. ML environment setup - you can use Vertex AI Workbench notebooks for development and experimentation. For more information you can check this link: https://cloud.google.com/architecture/ml-on-gcp-best-practices#machine-learning-environment-setup

2. ML Development - this focuses on data preparation, experimentation, and model evaluation. 
For more information you can check this link: https://cloud.google.com/architecture/ml-on-gcp-best-practices#machine-learning-development

3. Data processing - There are a lot of scenarios on how to process a data, you can check this link for the recommended approach on the common scenarios: https://cloud.google.com/architecture/ml-on-gcp-best-practices#data-processing

4. Operationalized training

Operationalized training refers to the process of making model training repeatable, tracking repetitions, and managing performance. 

---

For more information, you can check this link: https://cloud.google.com/architecture/ml-on-gcp-best-practices#operationalized-training

5. Model Deployment and serving this is where the model gets into production. Fore more information: https://cloud.google.com/architecture/ml-on-gcp-best-practices#model-deployment-and-serving

6. ML Workflow Orchestration - This is where you automate your Machine Learning workflows using Vertex AI pipelines that will allow you to retrain your models if necessary. For more information: https://cloud.google.com/architecture/ml-on-gcp-best-practices#machine-learning-workflow-orchestrati...

7. Artifact Organization - standardizing the artifacts from your Machine Learning workflow. For more information: https://cloud.google.com/architecture/ml-on-gcp-best-practices#artifact-organization

8. Model Monitoring - This is where you will monitor your model if it is performing as intended or if it is not performing as expected. For more information: https://cloud.google.com/architecture/ml-on-gcp-best-practices#model-monitoring

You can check this best practices for implementing machine learning on Google Cloud, this will help you understand the scope of Machine Learning in GCP and can help you plan accordingly: https://cloud.google.com/architecture/ml-on-gcp-best-practices

For your token and character limitations, you can try dividing your data into smaller parts that each contains a complete context in this way the model may be able to understand the contexts and the model can come up with a suitable response.  

You can also check these Blogposts in Google Research, it may be able to provide some key points for your inquiry.
https://ai.googleblog.com/2023/05/larger-language-models-do-in-context.html
https://ai.googleblog.com/2021/12/more-efficient-in-context-learning-with.html

Hope this helps!