Low-Code GPT-4 Workflow for Indonesian Sentiment Analysis with Dify

Abstrak

Businesses are increasingly seeking insights from textual data. However, traditional sentiment analysis pipelines require significant data preparation, model training, and the use of multiple tools. This paper presents a low-code, end-to-end sentiment analysis workflow. It is built on Dify and GPT-4, using few-shot prompting without fine-tuning. The system ingests batch text from spreadsheets, invokes GPT-4 with structured JSON outputs (including sentiment score, label, categories, and keywords), and then exports results for downstream use. It comprises three layers: scraping Indonesian protest tweets, AI inference via Dify, and interactive visualization through “Sentiment Dashboard Lite.” We evaluate the framework on 353 protest-related tweets (20 Aug–5 Sep 2025), comparing GPT-4-turbo (multi-sentiment), GPT-4o-mini (single-sentiment), and a traditional IndoBERT classifier. The two GPT-4 variants show moderate agreement (κ ≈ 0.547) but diverge from IndoBERT (κ ≈ 0.165–0.181), reflecting different labeling behavior. GPT-4-turbo processes more tokens (444 vs. 417), with higher latency (3.71 vs. 1.97s) and cost (US$5.80 vs. US$0.09 per 1,000 texts), but both yield throughput over 900 texts/hour. This case study demonstrates that a low-code GPT-4 pipeline can deliver rich, structured sentiment insights on Indonesian text with minimal engineering, supporting agile, data-driven decisions.

Identifier

• OpenAlex: https://openalex.org/W7125377257
• DOI: https://doi.org/10.1145/3786554.3786574