GenAI End to End

The intent of this blog is to present the a very high-level explanation of the end-to-end processing involved in a typical interaction with the Gen AI model:

• User: The process begins with the user, who initiates the AI interaction. This usually happens over a chatbot console like ChatGPT.
• Generate Prompt: The user generates a prompt or input they want the AI to process. A prompt is generally a text statement in users' natural language called NLP in the AI/ML context.
• Pre-Processing: Before tokenisation, there might be a pre-processing step where the input is cleaned or formatted. Sometimes, this may involve adding steps to check sensitivity and bias in the context.
• Tokenisation: The input is tokenised into manageable pieces, like words or sub-words.
• Generate Embedding: The tokens are converted into embeddings, vector representations that capture the semantic meaning.
• Positional Encoding: Positional encodings are added to embeddings to give the model information about the order of the tokens. These steps are part of the AI service processing.
• Core Transformer Architecture: The embeddings are passed into the core Transformer model, which consists of an encoder and decoder.
• Attention Layer (Encoder): In the encoder, the attention mechanism helps the model to focus on different parts of the input sequence.
• Feed-Forward Networks: Inside the encoder and decoder are feed-forward networks in addition to the attention layers.
• Multi-Head Attention: The attention mechanism is often multi-headed, meaning it runs several attention processes in parallel.
• Attention Layer (Decoder): The decoder's attention mechanism helps generate the output by focusing on relevant and partially generated parts of the input.
• Output Embedding: The decoder's output is transformed into an output embedding.
• Softmax Layer: A softmax layer converts the output embeddings into probabilities for the following tokens.
• Post-Processing: The output may be post-processed to refine it, such as correcting grammar or style.
• Output: The processed output is delivered back to the user.

Additional Considerations:

• Layer Normalization and Residual Connections: These are included within each attention and feed-forward layer to help train deeper models.
• Training Process: The model adjusts its parameters based on a large dataset during the training phase.
• Parameter Optimization: The model parameters are adjusted through optimisation techniques to minimise the loss function.
• Loss Calculation: The model calculates the loss to measure the difference between its predictions and the actual outputs during training.
• Beam Search or Sampling: The model may use techniques like beam search or sampling to generate the most likely following tokens during decoding.
• Hyperparameters and Model Configuration: The model's configuration, such as the number of layers, embedding sizes, and hyperparameters, is crucial for performance but not shown in the flowchart.