03 Response control

Prompt engineering - create few-shot prompt examples

Clean the response and check if the LLM correctly returned both the aspect and the sentiment

Prompt engineering - create the full prompt for each review with instructions and examples

Set up the OpenAI API key, authenticate, and select the LLM. Send the prompt to LLM to extract aspect and sentiment.

Import few-shot examples

Rebuild the prompt to instruct the model to return the result in JSON format.

Convert the JSON response into table, extract aspect and sentiment, and evaluate output consistency

03 Response Control - Aspect Based Sentiment Analysis

This workflow demonstrates how a few-shot prompt can be used to perform aspect-based sentiment analysis an how to validate the model's response. It showcases response control techniques by guiding the model with labelled examples and checking whether the predicted output follows the intended format (string or JSON).

The workflow processes product review data for laptops, combining review content with few-shot examples to generate aspect-specific prompts. It then sends these prompts to an LLM, which is expected to identify both the aspect and the corresponding sentiment. After the response is received, validation steps compare the output against the original prompt, checking for consistency.

Prompt engineering - create few-shot prompt examples and format them as JSON

Option 1: return output as a string

Option 2: return output as JSON