AI-Driven Welfare Scheme Recommendation Using Random Forest and RAG

We propose a scalable framework that integrates a multi-output Random Forest classifier with a retrieval-augmented generation module to recommend government welfare schemes to citizens based on their demographic profiles. The system first applies a preprocessing pipeline that normalizes raw input features—such as age, income, occupation, and caste—using a fuzzy matching algorithm to resolve lexical inconsistencies in categorical variables. A multi-label Random Forest ensemble, comprising hundreds of decision trees, then predicts eligibility probabilities across all available schemes simultaneously, and a calibrated confidence threshold selects a candidate subset of schemes. To ensure factual accuracy in the natural language explanations delivered to users, we incorporate a retrieval-augmented generation component. This module embeds verified scheme descriptions into a high-dimensional vector space, retrieves the most relevant document chunks for the candidate schemes using cosine similarity, and feeds both the retrieved context and the user’s original query into an instruction-tuned large language model. The classification stage thus acts as a computational filter that narrows the retrieval search space, thereby improving both system efficiency and response precision. The primary contribution of this work lies in the novel coupling of an ensemble-based eligibility predictor with a retrieval-constrained generative model, which prevents hallucinated outputs while remaining adaptable to large-scale, heterogeneous citizen data. Experimental evaluations on synthetic datasets, designed to mimic real-world public records, demonstrate that the framework achieves high precision in eligibility prediction and generates coherent, evidence-backed recommendations. This approach has significant implications for making complex social welfare systems more accessible to underserved populations.