PENGEMBANGAN CHATBOT REKOMENDASI INFLUENCER BERBASIS RETRIEVAL-AUGMENTED GENERATION PADA PLATFORM MANAJEMEN KAMPANYE DIGITAL PT MITRA OPTIMASI HANDAL

Yudhistira Rafazaky Bandono, Dr. Imam Fahrurrozi , S.T., M.Cs.

2026 | Tugas Akhir | D4 Teknologi Perangkat Lunak

Pengelolaan kampanye influencer marketing membutuhkan analisis data heterogen dalam jumlah besar, namun pendekatan dashboard statis yang ada dinilai kurang efektif dalam menghasilkan insight yang cepat dan kontekstual. Proyek Akhir ini mengembangkan chatbot rekomendasi influencer berbasis Retrieval-Augmented Generation (RAG) menggunakan LangGraph dengan model Qwen3.5:0.8B, dikombinasikan dengan pgvector dan strategi pencarian hybrid Reciprocal Rank Fusion (RRF), serta backend FastAPI dengan streaming Server-Sent Events pada platform manajemen kampanye digital PT Mitra Optimasi Handal. Pemilihan model dilakukan melalui evaluasi komparatif terhadap tiga Large Language Model (LLM) menggunakan 150 test case dari Stanford Question Answering Dataset (SQuAD), di mana Qwen3.5:0.8B unggul dengan nDCG@5 sebesar 0,9025 dan Recall@5 sempurna 1,0000. Hasil pengujian menunjukkan seluruh skenario black-box testing berjalan sesuai spesifikasi, load testing membuktikan stabilitas sistem dengan error rate 0,00?n rata-rata response time non-LLM di bawah 100 ms, serta User Acceptance Testing (UAT) memperoleh skor 90,85?ri tujuh responden. Sistem ini diharapkan membantu tim pemasaran dalam pengambilan keputusan pemilihan influencer secara lebih cepat, akurat, dan berbasis data. 

Managing influencer marketing campaigns requires analyzing large volumes of heterogeneous data, yet existing static dashboard approaches are insufficient for generating fast and contextual insights. This Final Project develops an influencer recommendation chatbot based on Retrieval-Augmented Generation (RAG) using LangGraph with the Qwen3.5:0.8B model, combined with pgvector storage and a Reciprocal Rank Fusion (RRF) hybrid retrieval strategy, and a FastAPI backend with Server-Sent Events streaming on PT Mitra Optimasi Handal's digital campaign management platform. Model selection was conducted through a comparative evaluation of three Large Language Models (LLMs) using 150 test cases from the Stanford Question Answering Dataset (SQuAD), where Qwen3.5:0.8B achieved the highest nDCG@5 of 0.9025 and a perfect Recall@5 of 1.0000. Testing results show all black-box testing scenarios passed specifications, load testing confirmed system stability with a 0.00% error rate and average non-LLM response time under 100 ms, and User Acceptance Testing (UAT) yielded a score of 90.85% from seven respondents. This system is expected to help marketing teams make influencer selection decisions more quickly, accurately, and in a data-driven manner.

Kata Kunci : Marketing, Large Language Model, Retrieval-Augmented Generation, Hybrid Retrieval, Reciprocal Rank Fusion, LangGraph, Qwen3.5:0.8B, pgvector, FastAPI, User Acceptance Testing