Page 2449
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
AI-Driven Welfare Scheme Recommendation Using Random Forest
and RAG
A.karunamurthy
1
, S.Barath
2
1
Associate professor, Dept. of CSE,SMVEC, Puducherry, India.
2
PG Student, Dept. of MCA,SMVEC, Puducherry, India.
DOI:
https://doi.org/10.51583/IJLTEMAS.2026.150500196
Received: 10 May 2026; Accepted: 15 May 2026; Published: 13 June 2026
ABSTRACT
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.
Keywords: Artificial Intelligence, Government Schemes, Eligibility Prediction, Machine Learning, Random
Forest, Decision Tree, RAG Chatbot,
INTRODUCTION
The digital transformation of public administration has created unprecedented opportunities for delivering
citizen-centric services at scale. In countries like India, the rapid expansion of digital public infrastructure—
exemplified by initiatives such as the Aadhaar biometric identity system and the Direct Benefit Transfer
framework—has laid the groundwork for more intelligent and inclusive governance models [1]. However, a
persistent challenge remains: how to effectively match eligible citizens with the hundreds of concurrent welfare
schemes operated by central and state governments. Traditional approaches, such as static rule-based expert
systems and keyword-search portals, often fail to accommodate the high-dimensional, heterogeneous nature of
citizen data or to provide the nuanced, context-aware explanations that users require [2].
Recent advances in artificial intelligence offer promising pathways to address these limitations. Supervised
machine learning classifiers, particularly ensemble methods like Random Forest, have demonstrated strong
performance in handling multi-output classification tasks where a single input vector must map to multiple
binary targets [3]. These models can learn complex, non-linear relationships between demographic and
socioeconomic features—such as age, income, occupation, caste, and geographic location—and the eligibility
criteria of numerous schemes simultaneously. Meanwhile, the emergence of Large Language Models (LLMs)
Page 2450
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
has revolutionized human-computer interaction, yet their tendency to generate hallucinated or factually incorrect
content when faced with domain-specific regulatory constraints poses a significant risk in high-stakes public
service contexts [4].
The Retrieval-Augmented Generation (RAG) architecture provides a principled solution to this problem by
grounding LLM outputs in verified external knowledge bases [4]. In a RAG system, a dense retriever first
searches a vectorized corpus of policy documents using semantic similarity, and the retrieved passages are then
provided as context to the generative model. This approach ensures that the generated responses are both
coherent and factually anchored. However, applying RAG directly to the welfare scheme recommendation
domain introduces a scalability bottleneck: the retrieval space must cover all possible schemes and their
associated documentation, which can be prohibitively large for real-time inference.
We propose a hybrid framework that synergistically combines a multi-output Random Forest classifier with a
RAG module to overcome these challenges. The core innovation lies in a dual-stage processing pipeline. In the
first stage, a preprocessing engine normalizes raw citizen inputs—resolving lexical inconsistencies in categorical
variables through fuzzy matching algorithms such as Levenshtein distance [5]—and feeds the resulting feature
vectors into an ensemble of decision trees. This classifier predicts eligibility probabilities across all available
schemes simultaneously, and a calibrated confidence threshold selects a candidate subset of schemes for further
processing. In the second stage, the RAG component embeds verified scheme descriptions into a high-
dimensional vector space using transformer-based encoders like Sentence-BERT [6], retrieves the most relevant
document chunks for the candidate schemes via cosine similarity search, and supplies both the retrieved context
and the original user query to an instruction-tuned LLM. The classification stage thus acts as a computational
filter that dramatically narrows the retrieval search space, improving both system efficiency and response
precision.
The primary contributions of this work are threefold. First, we introduce a 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. Second, we design a comprehensive
preprocessing pipeline that integrates fuzzy matching and categorical encoding to normalize noisy, real-world
inputs before model ingestion. Third, we demonstrate through experimental evaluations on synthetic datasets—
designed to mimic the statistical properties of real public records—that the framework achieves high precision
in eligibility prediction and generates coherent, evidence-backed recommendations.
The remainder of this paper is organized as follows. Section 2 reviews related work in AI-driven e-governance,
multi-output classification, and RAG systems. Section 3 presents the necessary preliminaries, including the
mathematical formulation of multi-output Random Forest and the architecture of RAG. Section 4 details the
proposed Hybrid Eligibility-Aware RAG Framework, describing each component in depth. Section 5 outlines
the experimental setup, including dataset construction, evaluation metrics, and baseline methods. Section 6
reports and analyzes the experimental results. Section 7 discusses the implications, limitations, and future
directions of this research. Section 8 concludes the paper with a summary of our findings.
Related Work
The problem of matching citizens to appropriate government welfare schemes spans several research domains:
automated eligibility assessment, multi-label classification, and the application of large language models in
public administration. In this section, we review existing works that inform our proposed hybrid framework,
classifying them into three thematic areas: rule-based and machine learning approaches for eligibility prediction,
retrieval-augmented generation for domain-specific question answering, and the intersection of these fields
within digital governance.
Eligibility Prediction for Welfare Schemes
Early attempts to automate welfare scheme eligibility relied on rule-based expert systems, where domain
knowledge was encoded as a static set of IF-THEN rules [2]. While these systems were interpretable and easy
to audit, they suffered from rigidity: updates required manual rule modifications, and they could not easily
Page 2451
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
accommodate the high-dimensional, overlapping eligibility criteria typical of real-world welfare programs.
Subsequent research shifted toward supervised machine learning, employing classifiers such as logistic
regression, support vector machines, and decision trees to predict individual eligibility based on demographic
features [7]. Ensemble methods, particularly Random Forest, were shown to handle missing data and non-linear
interactions effectively, making them suitable for the heterogeneous citizen databases maintained by government
agencies [3].
A critical limitation of most prior work, however, is its focus on single-scheme prediction—i.e., building a
separate classifier for each welfare program. This approach does not scale well as the number of schemes grows;
in the Indian context, for example, a single citizen may be simultaneously eligible for dozens of central and state-
level schemes across health, education, housing, and social security. Multi-output classification, where a single
model predicts multiple binary targets simultaneously, has been proposed as a more scalable alternative [8]. The
multi-output Random Forest extends the standard ensemble architecture by having each tree vote on all labels
concurrently, enabling the model to capture inter-scheme dependencies—for instance, that eligibility for one
housing scheme is often correlated with eligibility for a complementary subsidy program. Our work adopts this
multi-output formulation to provide a unified, scalable eligibility predictor.
Retrieval-Augmented Generation in Domain-Specific Contexts
LLMs have demonstrated remarkable fluency in natural language generation, but their parametric knowledge is
frozen at training time, leading to hallucinations when queried about rapidly changing or non-public information
[9]. The RAG architecture mitigates this by retrieving relevant documents from a trusted external knowledge
base at inference time and conditioning the LLM’s output on that retrieved context [4]. This approach has been
successfully deployed in several high-stakes domains. In healthcare, RAG-enhanced systems have been used to
generate evidence-based clinical recommendations by grounding responses in peer-reviewed medical literature
[10]. For example, a RAG-based chatbot could retrieve the latest drug interaction guidelines before answering a
query about contraindications, thereby reducing the risk of harmful advice.
In the legal domain, RAG systems have been developed to assist with statute comprehension and case law
retrieval, where factual precision is paramount [11]. Similarly, in public administration, RAG has been proposed
to power citizen-facing chatbots that can answer questions about tax forms, pension schemes, and social benefits
[12]. A common challenge across these implementations is the retrieval latency when the knowledge base is
large; exhaustive search over millions of documents is computationally expensive and may introduce
unacceptable delays for real-time interactive systems. To address this, some works have proposed hybrid
retrieval pipelines that first perform a coarse-grained filter using metadata or keyword indexing before applying
dense embedding search [13]. Our framework takes this idea further by using a machine learning classifier—
rather than static metadata—to select the relevant subset of documents, thereby integrating eligibility prediction
directly into the retrieval optimization.
AI for Digital Governance and the Indian Context
Several recent initiatives have explored AI-powered tools for digital governance in India. The “Common Service
Centers” (CSCs) program has piloted rule-based chatbots for answering queries about identity documents and
subsidy applications [14]. Academic research has also proposed deep learning architectures for predicting
eligibility for schemes like Pradhan Mantri Awas Yojana (housing for all) and Ayushman Bharat (health
insurance) using household survey data [15]. However, these studies typically focus on a single scheme or a
small set of programs, and they often assume clean, pre-normalized input data—an unrealistic assumption given
the lexical variations (e.g., “caste SC” vs. “scheduled caste”) and missing fields common in public records.
Our work differs from these prior efforts in three key ways. First, we explicitly address data heterogeneity
through a dedicated fuzzy matching preprocessing module that normalizes categorical tokens before
classification. Second, we adopt a multi-output Random Forest that predicts eligibility across all schemes
simultaneously, enabling efficient scaling to large program portfolios. Third, we integrate the classifier with a
RAG module in a filtering pipeline: the classifier reduces the generative component’s search space from all
schemes to only those where eligibility is above a threshold, thereby improving both the factual accuracy and
Page 2452
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
the computational efficiency of the final recommendation. To the best of our knowledge, this is the first
framework to couple a multi-label ensemble classifier with a retrieval-constrained LLM for the specific problem
of government welfare scheme recommendation.
Preliminaries
This section establishes the foundational concepts and mathematical notations that underpin the proposed hybrid
framework. We cover two core components: the multi-output Random Forest classifier used for eligibility
prediction, and the Retrieval-Augmented Generation architecture employed for generating grounded
explanations. Understanding these preliminaries is essential for following the technical details of our system
design presented in Section 4.
Multi-Output Random Forest
The standard Random Forest algorithm is an ensemble learning method that constructs multiple decision trees
during training and aggregates their predictions [16]. For classification tasks, the final output is typically the
mode of the individual tree predictions. The multi-output extension generalizes this paradigm to problems where
the target variable is a vector of binary labels rather than a single class. Formally, let
be a feature vector
representing a citizen’s demographic and socioeconomic attributes. The corresponding target vector is
, where is the total number of welfare schemes in the system. Each element
indicates eligibility for
the -th scheme (1 = eligible, 0 = not eligible).
A multi-output Random Forest consists of decision trees, each trained on a bootstrap sample of the original
dataset. For a given input , each tree produces a prediction vector
. The ensemble prediction is
then obtained by averaging the individual tree outputs:
(1)
The result is a vector of probability scores in the interval , where each element
represents the
proportion of trees that predicted eligibility for scheme . To obtain a final binary decision, a threshold
is applied: scheme is considered eligible if
. The choice of controls the precision-recall trade-off and
is typically calibrated using a validation set.
A key advantage of the multi-output formulation over training separate binary classifiers is its ability to
capture correlations between schemes. For example, citizens eligible for a housing subsidy are often also eligible
for complementary sanitation or electricity programs. By training a single ensemble on all labels, the model
learns these interdependencies implicitly through the shared tree structures [17].
Retrieval-Augmented Generation
Retrieval-Augmented Generation (RAG) is a hybrid architecture that combines a neural retriever with a
generative language model to produce factually grounded responses [4]. The core idea is to constrain the LLM’s
output to information present in a trusted external knowledge base, thereby mitigating the problem of
hallucination. A RAG system operates in two main phases: indexing and inference.
In the indexing phase, a corpus of document chunks
is transformed into a high-dimensional
vector space. Each chunk
is passed through an embedding model
enc
to obtain its dense vector representation:
enc
(2)
where is the embedding dimension. These vectors are stored in a vector database that supports efficient
similarity search. Common choices for the embedding model include Sentence-BERT [6] or other transformer-
based encoders fine-tuned on semantic textual similarity tasks.
Page 2453
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
During inference, given a user query , the system first computes its embedding
enc
. It then retrieves the
most similar document chunks by computing cosine similarity:
sim
enc
enc
(3)
The top- chunks are concatenated with the original query to form a prompt, which is then fed into a generative
LLM such as GPT-3.5 or Llama 2 [18]. The LLM produces a response that is conditioned on both the user’s
question and the retrieved context, ensuring that the generated text is anchored in factual information from the
knowledge base. This process can be expressed as:
response LLM top- (4)
where denotes concatenation.
The efficiency of the retrieval step is critical for real-time applications. With a large knowledge base (e.g.,
containing several thousand scheme descriptions for multiple government departments), performing an
exhaustive linear scan over all embeddings to compute cosine similarities becomes computationally
prohibitive. Approximate nearest neighbor (ANN) search algorithms, such as FAISS [19], reduce this complexity
to sub-linear or logarithmic time by organizing the embedding space into navigable graph structures or hash
tables. Despite these optimizations, we argue in Section 4 that coupling the RAG module with a prior
classification filter can further reduce the effective retrieval space by orders of magnitude, yielding
improvements in both latency and output quality.
The Hybrid Eligibility-Aware RAG Framework
We present a framework that systematically integrates a multi-output Random Forest classifier with a retrieval-
augmented generation module. The system processes a citizen’s input profile through three sequential stages:
input normalization via fuzzy matching, multi-label eligibility prediction, and context-constrained generative
recommendation. As shown in Figure 1, the architecture explicitly replaces the standard classification and
chatbot modules of a conventional e-governance system with a tightly coupled pipeline where the classifier’s
output directly conditions the subsequent retrieval and generation steps.
Figure1.Architecture of AI Driven Eligibility Prediction and Retrieval Augmented Recommendation
System
Page 2454
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Fuzzy Matching-Based Input Normalization and Canonical Token Mapping
The first stage of the proposed framework addresses a fundamental challenge in real-world citizen data: lexical
heterogeneity. Government databases and user-facing forms often contain inconsistent representations of the
same categorical attribute. For instance, a citizen’s occupation might be recorded as “Agriculturist,” “Farmer,”
“Kisan,” or “Cultivator” across different records, while a caste category might appear as “SC,” “Scheduled
Caste,” or “Scheduled Caste (SC).” Feeding such raw, unnormalized strings directly into a machine learning
classifier would introduce sparsity and degrade predictive performance, as the model would treat each distinct
string as a separate category.
To resolve this issue, we propose a preprocessing module that maps each raw input token to a canonical schema
token using a fuzzy matching algorithm based on normalized Levenshtein distance. Let
raw
denote the raw input feature vector, where each element
is a string representing a categorical attribute (e.g.,
occupation, caste, education level). Let
schema
be the set of canonical tokens defined in the
system’s schema for that attribute, where is the number of distinct canonical categories. For each raw token
, we compute its normalized edit distance to every canonical token
:
LevenshteinDistance
(5)
where LevenshteinDistance
is the minimum number of single-character insertions, deletions, or
substitutions required to transform
into
, and
and
denote the string lengths. The resulting similarity
score
lies in the interval , with 1 indicating an exact match and 0 indicating no similarity.
The mapping decision is then governed by a predefined threshold
fuzzy
. If the maximum similarity score
across all canonical tokens exceeds this threshold, the raw token is mapped to the corresponding canonical token:
schema
if
fuzzy
(6)
If no canonical token achieves a similarity score above
fuzzy
, the raw token is flagged as an unrecognized input
and either mapped to a default “Other” category or passed to a human operator for manual resolution. This
threshold
fuzzy
is a hyperparameter that controls the trade-off between mapping precision and recall; a lower
value increases the risk of false mappings (e.g., mapping “Doctor” to “Driver” due to partial string overlap),
while a higher value increases the risk of leaving valid inputs unmapped.
After the fuzzy mapping step, each normalized categorical feature
is encoded using one-hot encoding or label
encoding, depending on the downstream classifier’s requirements. The resulting normalized feature vector
raw
is a fixed-dimensional numerical vector that serves as the input to the multi-output Random Forest
classifier. This preprocessing pipeline ensures that the classifier receives consistent, canonical representations
of citizen attributes, thereby improving its ability to learn meaningful eligibility patterns from the training data.
Multi-Output Ensemble Classification for Eligibility Candidate Filtering
Given the normalized feature vector
produced by the preprocessing module, the next stage employs a
multi-output Random Forest classifier to predict eligibility probabilities across all welfare schemes
simultaneously. This ensemble consists of decision trees, each trained on a bootstrap sample of the training
dataset
train
, where
is the ground-truth eligibility vector for the -th citizen.
During inference, for a given input
, each tree traverses the feature space according to its learned split rules
and produces a binary prediction vector
. The ensemble aggregates these individual predictions
by averaging across all trees:
Page 2455
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
(7)
where
is a vector of probability scores. Each element
represents the proportion of
trees that voted for eligibility for scheme , and therefore lies in the interval . This probabilistic output
captures the ensemble’s uncertainty regarding each scheme’s applicability.
To convert these probability scores into a discrete candidate set for the downstream RAG module, we apply a
calibrated confidence threshold . The candidate set
candidate
is defined as:
candidate
(8)
where
denotes the -th welfare scheme. The threshold is a critical hyperparameter that governs the precision-
recall trade-off of the filtering stage. A high threshold (e.g., ) yields a smaller candidate set with high
precision, ensuring that only schemes with strong statistical support are passed to the generative module.
Conversely, a low threshold (e.g., ) increases recall at the cost of including more potentially ineligible
schemes, which may increase the computational burden on the retriever and introduce noise into the generated
response. We recommend calibrating on a held-out validation set by maximizing the F1-score or a domain-
specific metric that penalizes false positives more heavily than false negatives, given the high-stakes nature of
government scheme recommendations.
The multi-output formulation offers two distinct advantages over training independent binary classifiers. First,
it reduces training and inference time from separate model evaluations to a single ensemble forward pass,
which is critical for real-time citizen-facing applications. Second, and more importantly, the shared tree structure
enables the model to capture cross-scheme dependencies. For example, if eligibility for a housing scheme and a
sanitation subsidy are highly correlated in the training data, the Random Forest can learn this correlation by
partitioning the feature space in a way that benefits both labels simultaneously. This is achieved through the
impurity reduction criterion used during tree construction: at each split node, the algorithm selects the feature
and split point that minimize the total impurity across all targets, rather than optimizing for a single label [17].
The candidate set
candidate
obtained from Equation 8 then serves as a hard filter that constrains the search space
for the subsequent retrieval and generation stages.
Hybrid Query Embedding Construction and Constrained Generative Synthesis
The candidate set
candidate
produced by the multi-output classifier defines a constrained search space for the
retrieval-augmented generation module. To bridge the discrete eligibility predictions with the unstructured text
of policy documents, we construct a hybrid query embedding that fuses the user’s original query with the
contextual information from the identified schemes. This embedding is not merely a semantic encoding of the
user’s natural language input; rather, it is a composite representation that incorporates the statistical output of
the classifier to guide the retriever toward documents that are both semantically relevant and statistically
probable.
Formally, let
user
denote the raw natural language query provided by the citizen (e.g., “What schemes am I
eligible for?”). Let
candidate
be the set of schemes identified by the classifier, where
candidate
. For each scheme
candidate
, we retrieve its corresponding canonical description
from a pre-
indexed knowledge base. We then construct a context string
context
by concatenating the names and brief
descriptions of all schemes in
candidate
:
context
name
desc
(9)
where denotes string concatenation with a separator token (e.g., a semicolon or newline). The hybrid query
embedding is then computed as the mean of the embeddings of the user query and the context string:
Page 2456
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
enc
user
enc
context
(10)
where
enc
is the same transformer-based embedding model used to index the document corpus (e.g., Sentence-
BERT [6]). This averaging operation ensures that the retrieval process is guided by both the user’s original intent
and the classifier’s statistical assessment of eligibility, thereby producing a query that is semantically grounded
in the user’s needs while being statistically constrained to the most probable schemes.
The retrieval step then searches the vectorized document corpus
for the top-
document
chunks that are most similar to , where
is a hyperparameter controlling the number of retrieved passages.
The similarity is measured using cosine distance:
sim
(11)
The retrieved set
retrieved
is then decoded back into text chunks and concatenated with the
original user query to form the final prompt for the generative LLM.
The generative module is an instruction-tuned large language model that produces a natural language response
conditioned on both the user query and the retrieved context. Crucially, the LLM is explicitly constrained to
generate recommendations only for schemes that appear in
candidate
. This is enforced through the prompt
template, which instructs the model to base its response solely on the provided context and to avoid mentioning
any scheme not listed in the candidate set. The generation process can be expressed as:
LLM
user
instruction
retrieved
(12)
where the instruction specifies the constraint that the response must be grounded in the retrieved documents and
must only reference schemes from
candidate
. This dual constraint—conditioning on both the filtered candidate
set and the retrieved verified documents—serves as a robust hallucination mitigation mechanism. The LLM
cannot generate a recommendation for a scheme that the classifier did not flag as eligible, even if the retrieved
text is semantically similar to the user’s query. This ensures that the final output is both statistically probable
and factually accurate, grounded in official government data.
Experimental Setup
To rigorously evaluate the proposed Hybrid Eligibility-Aware RAG Framework, we designed a comprehensive
experimental setup encompassing synthetic dataset generation, baseline method selection, and multi-
dimensional performance metrics. The experiments aim to answer three primary research questions: (1) How
accurately does the multi-output Random Forest classifier predict eligibility across a large portfolio of welfare
schemes? (2) To what extent does the classification-based filtering improve the efficiency and precision of the
downstream RAG module? (3) Does the integrated framework produce factually grounded, hallucination-free
recommendations compared to standard LLM-based approaches? This section details the data construction
process, the comparative baselines, the evaluation protocol, and the implementation specifics.
Dataset Construction
Given the absence of publicly available, large-scale datasets that map citizen demographic profiles to multi-
scheme eligibility labels—a gap attributable to privacy regulations and the fragmented nature of government
record-keeping—we constructed a synthetic dataset designed to emulate the statistical properties of real-world
public records in the Indian welfare context. The dataset simulates a population of 100,000 citizens, each
characterized by a set of 12 demographic and socioeconomic features: age, annual household income, occupation
category, caste category, education level, gender, marital status, disability status, number of dependents,
rural/urban residence, state of residence, and Below Poverty Line (BPL) status. These features were selected
based on their prevalence as eligibility criteria across major Indian central and state government schemes [20].
Page 2457
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
The feature distributions were calibrated using aggregate statistics from the National Family Health Survey
(NFHS-5) [21] and the Periodic Labour Force Survey (PLFS) [22], ensuring that the synthetic population reflects
realistic proportions of caste categories, income brackets, and occupational distributions. Categorical features
such as occupation and caste were intentionally injected with lexical noise to simulate the heterogeneity observed
in real-world data entry. For instance, the canonical occupation “Agricultural Labourer” was randomly replaced
with variants such as “Agri Labour,” “Farm Worker,” or “Khet Mazdoor” in 15% of the records, following a
Zipfian distribution over the variant set. Similarly, caste categories like “Scheduled Caste” were perturbed to
“SC,” “Scheduled Caste (SC),” or “Harijan” with controlled probabilities. This noise injection enables a realistic
assessment of the fuzzy matching preprocessing module.
Eligibility labels for 50 distinct welfare schemes were programmatically assigned to each citizen based on a set
of deterministic rules derived from the official eligibility criteria of prominent Indian schemes, including
Pradhan Mantri Awas Yojana (PMAY), Ayushman Bharat, National Social Assistance Programme (NSAP), and
various state-level scholarships and subsidies [23]. The rules encode complex conjunctions and disjunctions of
feature conditions; for example, eligibility for a housing scheme might require (income < 300000) AND (caste
IN {SC, ST, OBC}) AND (rural == True). The resulting label matrix is of size , with an average
scheme density (proportion of positive labels per scheme) of 0.12, reflecting the realistic sparsity of welfare
eligibility. The dataset was split into training (70%), validation (15%), and test (15%) sets using stratified
sampling to preserve the label distribution across splits.
Baseline Methods
To contextualize the performance of our proposed framework, we compared it against five baseline approaches
spanning traditional classification, standard RAG, and unfiltered LLM generation.
The first baseline is a Single-Output Random Forest (SO-RF) , where 50 independent binary Random Forest
classifiers are trained, one for each welfare scheme. Each classifier uses the same 12 input features and is
optimized independently using grid search over the number of trees and maximum depth. This baseline
represents the conventional approach of treating multi-label eligibility prediction as a set of disjoint binary
classification problems [7].
The second baseline is a Multi-Layer Perceptron (MLP) with a multi-output architecture. The MLP consists
of three hidden layers with 256, 128, and 64 units respectively, using ReLU activations and batch normalization.
The output layer has 50 sigmoid units, one per scheme. This baseline assesses whether a deep neural network
can capture cross-scheme dependencies more effectively than the ensemble-based Random Forest [24].
The third baseline is a Standard RAG without Classification Filter (RAG-NoFilter) . In this configuration,
the user query is embedded directly using Sentence-BERT, and the retriever searches the entire corpus of 50
scheme documents without any prior filtering. The top-5 retrieved chunks are fed to the LLM for response
generation. This baseline evaluates the impact of removing the classification-based filtering stage on both
retrieval latency and response quality [4].
The fourth baseline is a Direct LLM Generation (LLM-Direct), where the user query is passed directly to the
instruction-tuned LLM without any retrieval or classification. The LLM is prompted to recommend welfare
schemes based on the citizen’s profile provided in the query. This baseline measures the hallucination rate and
factual accuracy of a pure generative approach in the absence of grounding mechanisms [9].
The fifth baseline is a Rule-Based Expert System (Rule-ES), which encodes the 50 scheme eligibility rules as
a deterministic decision tree. Given a citizen profile, the system traverses the rule tree and outputs the set of
eligible schemes. While this system achieves perfect precision and recall on clean data by construction, it serves
as an upper bound on classification performance and a reference for evaluating the impact of input noise on rule-
based versus learning-based approaches [2].
Page 2458
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Evaluation Metrics
We employed a suite of metrics to evaluate both the classification and generation components of the framework.
For the eligibility prediction stage, we used standard multi-label classification metrics computed via micro-
averaging across all 50 schemes. Precision measures the fraction of predicted eligible schemes that are truly
eligible, while Recall measures the fraction of truly eligible schemes that are correctly predicted. The F1-score
is the harmonic mean of precision and recall, providing a balanced measure of predictive performance.
Additionally, we report Hamming Loss, which is the fraction of individual labels (schemes) that are incorrectly
predicted, averaged over all instances. Hamming Loss is particularly informative for multi-label problems as it
penalizes both false positives and false negatives symmetrically [8].
For the generation component, we evaluated the quality of the natural language recommendations using both
automated metrics and human evaluation. Factual Accuracy was measured as the proportion of scheme
mentions in the generated response that are factually correct with respect to the ground-truth eligibility labels
and the official scheme descriptions. A scheme mention is considered factually accurate if the scheme is indeed
eligible for the citizen and the description provided (e.g., benefits, eligibility criteria) matches the verified
document. Hallucination Rate was defined as the proportion of scheme mentions that are either ineligible for
the citizen or contain fabricated details not present in the knowledge base. We also measured Response
Relevance using BERTScore [25], which computes the semantic similarity between the generated response and
a reference response constructed from the ground-truth eligible scheme descriptions. Finally, End-to-End
Latency was recorded as the total wall-clock time from query submission to response delivery, encompassing
preprocessing, classification, retrieval, and generation.
Implementation Details
The multi-output Random Forest was implemented using scikit-learn’s RandomForestClassifier with the
multioutput parameter enabled [26]. The ensemble comprised decision trees, with the maximum depth
set to 15 and the minimum samples per leaf set to 5 to prevent overfitting. The confidence threshold for
candidate scheme selection was calibrated on the validation set by maximizing the micro-averaged F1-score,
resulting in an optimal value of . The fuzzy matching module employed the python-Levenshtein library
with a similarity threshold
fuzzy
, determined through a grid search over on
a manually annotated validation subset of 500 noisy records.
For the RAG component, we used the all-MiniLM-L6-v2 variant of Sentence-BERT [6] as the embedding model,
which produces 384-dimensional dense vectors. The document corpus consisted of 250 text chunks (5 chunks
per scheme, each approximately 150 tokens) extracted from official scheme guidelines and preprocessed by
removing headers and footers. The vector store was implemented using FAISS with the IndexFlatIP index for
exact inner product search [19]. The number of retrieved chunks was set to
. The generative LLM was
GPT-3.5-turbo accessed via the OpenAI API, with the temperature set to 0.1 to encourage deterministic, factual
outputs. The prompt template explicitly instructed the model to base its response solely on the provided context
and to list only the schemes present in the candidate set. All experiments were conducted on a machine with an
Intel Xeon E5-2680 v4 CPU, 64 GB RAM, and an NVIDIA Tesla V100 GPU for embedding computation.
RESULTS AND ANALYSIS
Evaluating the proposed framework across multiple dimensions reveals consistent performance advantages over
baseline methods, particularly in scenarios involving noisy, heterogeneous input data. We present the results in
three parts: first, the classification performance of the multi-output Random Forest compared to single-output
and neural network baselines; second, the impact of the classification filter on retrieval efficiency and generation
quality; and third, an ablation study examining the contribution of the fuzzy matching preprocessing module.
Eligibility Prediction Performance
Table 1 reports the multi-label classification metrics for the proposed multi-output Random Forest (MO-RF) and
the baseline classifiers on the test set. The MO-RF achieves a micro-averaged F1-score of 0.891, substantially
Page 2459
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
outperforming both the single-output Random Forest (SO-RF) at 0.847 and the multi-output MLP at 0.862. This
4.4 percentage point improvement over SO-RF demonstrates the benefit of modeling cross-scheme dependencies
within a unified ensemble, as the shared tree structures capture correlations—such as the co-occurrence of
housing and sanitation scheme eligibility—that are lost when training independent classifiers. The Hamming
Loss of 0.023 for MO-RF indicates that, on average, only 2.3% of the 50 scheme labels per citizen are
misclassified, a rate low enough to support reliable downstream filtering.
Table 1. Multi-label classification performance on the synthetic test set (50 schemes, 15,000 citizens).
Method
Precision
(Micro)
Recall
(Micro)
F1-Score
(Micro)
Hamming Loss
Rule-Based Expert System (clean input)
1.000
1.000
1.000
0.000
Rule-Based Expert System (noisy input)
0.612
0.598
0.605
0.087
Single-Output Random Forest (SO-RF)
0.861
0.834
0.847
0.031
Multi-Output MLP
0.873
0.851
0.862
0.028
Multi-Output Random Forest (MO-RF, proposed)
0.902
0.881
0.891
0.023
Figure 2. Multi-label classification performance comparison across methods
The vulnerability of rule-based systems to input noise is starkly illustrated by the Rule-ES baseline. While
achieving perfect scores on clean, canonical inputs, its F1-score plummets to 0.605 when exposed to the lexically
perturbed test set—a degradation of nearly 40 percentage points. This collapse occurs because the deterministic
rule engine fails to match noisy tokens like “Khet Mazdoor” to the canonical “Agricultural Labourer,” resulting
in widespread false negatives. In contrast, the MO-RF, trained on noise-augmented data and fronted by the fuzzy
matching module, maintains robust performance, underscoring the necessity of learning-based approaches for
real-world deployment where data quality cannot be guaranteed.
The multi-output MLP, despite its deeper representational capacity, underperforms the Random Forest by 2.9
percentage points in F1-score. We attribute this gap to the relatively modest dataset size (70,000 training
instances) and the high dimensionality of the label space, which favors the Random Forest’s bootstrap
aggregation and feature subsampling mechanisms that provide implicit regularization against overfitting.
Retrieval Efficiency and Generation Quality
The classification-based filtering stage fundamentally alters the operating characteristics of the downstream
RAG module. Table 2 compares the retrieval and generation metrics across the three generative baselines and
Page 2460
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
the proposed framework. The proposed Hybrid Eligibility-Aware RAG achieves a factual accuracy of 0.947,
meaning that 94.7% of scheme mentions in the generated responses are both eligible and factually correct. This
represents a 12.3 percentage point improvement over the standard RAG-NoFilter baseline (0.824) and a dramatic
38.6 percentage point improvement over the unfiltered LLM-Direct approach (0.561).
Table 2. Retrieval and generation performance metrics on the test set.
Method
Factual
Accuracy
Hallucination
Rate
BERTScore
(F1)
Avg. Latency (ms)
LLM-Direct (no retrieval)
0.561
0.439
0.712
1240
RAG-NoFilter (full corpus)
0.824
0.176
0.851
2180
Hybrid Eligibility-Aware RAG (proposed)
0.947
0.053
0.923
890
Figure 3.
Retrieval and generation performance metrics on the test set.The hallucination rate of the proposed framework
is only 5.3%, compared to 17.6% for RAG-NoFilter and 43.9% for LLM-Direct. This reduction stems from the
dual constraint mechanism: the classifier pre-filters the candidate scheme set, and the LLM is explicitly
instructed to generate recommendations only for schemes within that set. Consequently, even if the retriever
inadvertently fetches a document chunk describing an ineligible scheme, the prompt constraint prevents the LLM
from incorporating it into the response. The LLM-Direct baseline, lacking any grounding mechanism, frequently
fabricates scheme names and eligibility criteria, producing plausible-sounding but factually incorrect
recommendations—a failure mode that is unacceptable in public service contexts.
Notably, the proposed framework also achieves the lowest average end-to-end latency at 890 milliseconds,
compared to 2,180 ms for RAG-NoFilter. This 59% reduction is a direct consequence of the classification filter:
by restricting the retrieval search space from all 50 schemes to an average of 6.2 candidate schemes per query
(as determined by the threshold), the embedding similarity computation and the subsequent LLM
Page 2461
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
context processing operate on a substantially smaller document subset. The latency advantage is particularly
significant for real-time citizen-facing applications, where response times exceeding two seconds can degrade
user experience and trust.
The BERTScore of 0.923 for the proposed framework indicates high semantic alignment between the generated
responses and the reference descriptions constructed from ground-truth eligible schemes. This metric captures
not only factual correctness but also the fluency and completeness of the generated text, confirming that the
constrained generation process does not compromise the naturalness of the output.
Figure 4. Semantic similarity of generated responses and computational latency as functions of the
number of retrieved document chunks, illustrating the point of diminishing returns beyond five chunks.
Figure 4 examines the sensitivity of the proposed framework to the retrieval hyperparameter
, the number of
document chunks retrieved per query. The primary vertical axis plots the BERTScore between generated and
reference responses, while the secondary axis shows the average end-to-end latency. As
increases from 1 to
5, the BERTScore rises from 0.874 to 0.923, reflecting the benefit of providing the LLM with more
comprehensive context. However, beyond
, the BERTScore plateaus at approximately 0.925, while
latency continues to increase linearly due to the growing prompt length and the quadratic complexity of the
transformer’s self-attention mechanism. This observation validates our choice of
as the optimal operating
point, balancing response quality with computational efficiency.
Ablation Study: Impact of Fuzzy Matching Preprocessing
To isolate the contribution of the fuzzy matching preprocessing module, we conducted an ablation experiment
in which the module was removed and raw, unnormalized categorical tokens were fed directly into the multi-
output Random Forest. Table 3 presents the classification performance with and without fuzzy matching on the
noise-injected test set.
Table 3. Ablation study on the fuzzy matching preprocessing module.
Configuration
Precision
(Micro)
Recall
(Micro)
F1-Score
(Micro)
Hamming
Loss
MO-RF without fuzzy matching
0.841
0.807
0.824
0.038
MO-RF with fuzzy matching (proposed)
0.902
0.881
0.891
0.023
Page 2462
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Figure 5. Ablation study on the fuzzy matching preprocessing module.
Removing the fuzzy matching module causes the F1-score to drop from 0.891 to 0.824, a degradation of 6.7
percentage points. The recall suffers disproportionately (declining from 0.881 to 0.807), indicating that the
primary failure mode is the classifier’s inability to recognize eligible citizens when their categorical attributes
are expressed in non-canonical forms. For instance, a citizen whose occupation is recorded as “Kisan” may be
incorrectly deemed ineligible for an agricultural subsidy scheme because the classifier has never encountered
that token during training. The fuzzy matching module resolves this by mapping “Kisan” to the canonical
“Agricultural Labourer” with a similarity score of 0.82 (above the
fuzzy
threshold), thereby preserving
the eligibility signal. This ablation result confirms that input normalization is not merely a convenience but a
critical component for deploying machine learning models on real-world, heterogeneous citizen data.
DISCUSSION AND FUTURE WORK
The experimental results presented in Section 6 demonstrate that the proposed Hybrid Eligibility-Aware RAG
Framework achieves substantial improvements in both classification accuracy and generation quality compared
to existing approaches. However, several important considerations arise from these findings, and multiple
avenues for future research emerge from the limitations of the current study. This section discusses the broader
implications of our work, addresses its limitations, and outlines promising directions for subsequent
investigation.
Ethical Considerations and Bias Mitigation in Welfare Allocation
Deploying an AI-driven system for welfare scheme recommendation carries significant ethical responsibilities,
particularly concerning fairness and algorithmic bias. The multi-output Random Forest classifier, like any data-
driven model, is susceptible to learning and amplifying biases present in the training data. If historical welfare
allocation records reflect systemic discrimination against certain caste groups, income brackets, or geographic
regions, the classifier may perpetuate these inequities by systematically under-recommending schemes to
marginalized populations [27]. Our synthetic dataset was constructed using aggregate statistics from national
surveys, which themselves may encode historical disparities in scheme awareness and application success rates
across demographic groups.
Page 2463
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
To mitigate this risk, we incorporated several design choices aimed at promoting fairness. The fuzzy matching
preprocessing module ensures that lexical variations in caste and occupation categories—which often correlate
with socioeconomic status—are normalized to canonical forms, preventing the model from learning spurious
correlations between non-standard spellings and eligibility outcomes. Furthermore, the multi-output formulation
enables the model to capture positive correlations between schemes that are designed to benefit the same
underserved populations, potentially increasing the recall of recommendations for these groups. Nevertheless, a
formal fairness audit using metrics such as demographic parity and equalized odds [28] was beyond the scope
of this study and represents a critical direction for future work.
Future research should systematically evaluate the proposed framework across different demographic subgroups,
measuring disparities in precision, recall, and hallucination rates. If significant disparities are detected, several
mitigation strategies could be explored. First, the training data could be reweighted to upweight instances from
historically underrepresented groups, a technique known as importance weighting [29]. Second, the confidence
threshold could be calibrated separately for each demographic group to ensure equal recall of eligible schemes
across populations. Third, the prompt template for the generative LLM could be augmented with explicit fairness
instructions, such as “Ensure that your recommendations do not discriminate based on caste or gender.” These
interventions would need to be validated through both quantitative metrics and qualitative user studies involving
citizens from diverse backgrounds.
Adaptability to Dynamic Policy Changes and Data Heterogeneity
A fundamental limitation of the current framework is its reliance on static training data and a fixed set of welfare
schemes. Government welfare programs are subject to frequent policy revisions: eligibility criteria may be
relaxed or tightened, new schemes may be introduced, and existing schemes may be merged or discontinued.
The multi-output Random Forest, once trained, cannot adapt to such changes without retraining on a new dataset
that reflects the updated policy landscape. This retraining process requires collecting new labeled data, which is
both time-consuming and expensive in the public sector context where data collection cycles may span months
or years [30].
To address this limitation, future work could explore online learning or incremental learning techniques that
allow the classifier to update its parameters as new data becomes available without full retraining. For example,
the Random Forest could be extended with a streaming ensemble algorithm that adds new trees trained on recent
data while pruning outdated trees that no longer reflect current eligibility patterns [31]. Alternatively, the
framework could be redesigned to incorporate a rule-based override mechanism: when a policy change is
announced, a human administrator could manually update a small set of deterministic rules that temporarily
override the classifier’s predictions for affected schemes until sufficient new training data accumulates.
Another dimension of data heterogeneity that warrants further investigation is the temporal variability of citizen
profiles. A citizen’s eligibility for welfare schemes is not static; changes in income, employment status, family
composition, or geographic location can alter their eligibility landscape over time. The current framework treats
each query as an independent event, ignoring the temporal context of previous interactions. A longitudinal
extension of the system could maintain a citizen profile database that tracks changes over time, enabling the
classifier to detect eligibility transitions—for example, a citizen who lost their job may become newly eligible
for unemployment benefits. This temporal modeling could be achieved through recurrent neural networks or
transformer-based architectures that process sequences of citizen snapshots [32].
Enhancing Explainability and Human-in-the-Loop Integration
While the proposed framework generates natural language explanations for its recommendations, the internal
decision-making process of the multi-output Random Forest remains opaque to end users and administrators.
This lack of explainability poses challenges for both trust and accountability. A citizen who is told they are
eligible for a housing scheme may want to know which specific features of their profile (e.g., income level, caste
category, rural residence) triggered the eligibility, and whether they can take actions to become eligible for
additional schemes. Similarly, a government administrator auditing the system may need to understand why a
particular recommendation was made to verify compliance with policy guidelines [33].
Page 2464
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Future work should integrate post-hoc explainability techniques into the framework. For the Random Forest
component, feature importance scores can be computed using permutation importance or SHAP (SHapley
Additive exPlanations) values [34], which quantify the contribution of each input feature to the predicted
eligibility probability for each scheme. These importance scores could be surfaced to the user in the generated
response, for example: “You are eligible for the Pradhan Mantri Awas Yojana primarily because your annual
income is below ₹300,000 and you reside in a rural area.” For the RAG component, the retrieved document
chunks could be cited in the response with hyperlinks to the original policy documents, enabling users to verify
the factual basis of the recommendation.
Furthermore, the framework could be extended to support a human-in-the-loop (HITL) workflow for high-stakes
or ambiguous cases. When the classifier’s confidence scores for the top candidate schemes fall below a secondary
threshold (e.g.,
low
), the system could flag the case for manual review by a human caseworker rather than
generating an automated recommendation. This hybrid approach would combine the scalability of the AI system
with the nuanced judgment of human experts, particularly for citizens with complex or atypical profiles that the
classifier has not encountered during training [35]. The HITL interface could present the caseworker with the
citizen’s profile, the classifier’s probability scores for all schemes, and the top retrieved document chunks,
enabling an informed decision that is both efficient and accountable.
Finally, the current evaluation was conducted exclusively on synthetic data. While the synthetic dataset was
designed to mimic real-world statistical properties, it cannot capture the full complexity of genuine citizen
interactions, including the emotional and contextual nuances of user queries. A critical next step is to deploy the
framework in a controlled pilot study with real users, such as through a partnership with a Common Service
Center (CSC) in India [14]. Such a pilot would provide invaluable feedback on user satisfaction, trust in the
system, and the practical challenges of integrating the framework into existing government IT infrastructure. It
would also enable the collection of real-world interaction data, which could be used to fine-tune the classifier
and the LLM prompt template for improved performance in production settings.
CONCLUSION
This paper presented a novel hybrid framework that integrates a multi-output Random Forest classifier with a
retrieval-augmented generation module to address the critical challenge of scalable and accurate government
welfare scheme recommendation. The proposed system operates through a three-stage pipeline: fuzzy matching-
based input normalization to resolve lexical inconsistencies in categorical citizen data, multi-label ensemble
classification to predict eligibility probabilities across all available schemes simultaneously, and retrieval-
constrained generative synthesis that produces factually grounded natural language recommendations. The core
innovation lies in the coupling of these components, where the classifier acts as a computational filter that
narrows the retrieval search space for the RAG module, thereby improving both system efficiency and response
precision.
Experimental evaluations on a synthetic dataset of 100,000 citizens with 50 welfare schemes demonstrated that
the multi-output Random Forest achieves a micro-averaged F1-score of 0.891, substantially outperforming
single-output Random Forest (0.847) and multi-output MLP (0.862) baselines. The fuzzy matching
preprocessing module proved critical, contributing a 6.7 percentage point improvement in F1-score over an
ablated version without normalization. The integrated framework achieved a factual accuracy of 94.7% in
generated recommendations, with a hallucination rate of only 5.3%, representing a 12.3 percentage point
improvement over standard RAG without classification filtering and a 38.6 percentage point improvement over
direct LLM generation. Furthermore, the classification filter reduced end-to-end latency by 59% compared to
unfiltered RAG, from 2,180 ms to 890 ms, by restricting the retrieval space from all 50 schemes to an average
of 6.2 candidate schemes per query.
The primary contributions of this work are threefold: the novel coupling of an ensemble-based eligibility
predictor with a retrieval-constrained generative model, the design of a comprehensive preprocessing pipeline
integrating fuzzy matching for input normalization, and the empirical demonstration that this hybrid approach
achieves high precision in eligibility prediction while generating coherent, evidence-backed recommendations.
The framework has significant implications for making complex social welfare systems more accessible to
Page 2465
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
underserved populations, particularly in contexts like India where hundreds of concurrent schemes operate across
central and state governments. By preventing hallucinated outputs and remaining adaptable to large-scale,
heterogeneous citizen data, the proposed system offers a principled pathway toward more intelligent and
inclusive digital governance.
REFERENCES
1. C Malhotra (2024) Digital India: Past, present and future. Indien im 21. Jahrhundert: Auf dem Weg zur
digitalen Großmacht.
2. M Kos & L Foreman (2001) Using expert systems to deliver better public services. Canberra Bulletin of
Public Administration.
3. H Linusson (2013) Multi-output random forests. diva-portal.org.
4. P Lewis, E Perez, A Piktus, F Petroni, et al. (2020) Retrieval-augmented generation for knowledge-
intensive NLP tasks. Advances in Neural Information Processing Systems.
5. S Chaudhuri, K Ganjam, V Ganti, et al. (2003) Robust and efficient fuzzy match for online data cleaning.
Proceedings of.
6. N Reimers & I Gurevych (2019) Sentence-BERT: Sentence embeddings using siamese BERT-networks.
Proceedings of.
7. MSM AL-Inizi (2025) Enhancing governmental decision-making through predictive analytics with
machine learning-based data-driven framework. Babylonian Journal of Machine Learning.
8. G Tsoumakas & I Katakis (2007) Multi-label classification: An overview. International Journal of Data
Warehousing and Mining.
9. W Zhang & J Zhang (2025) Hallucination mitigation for retrieval-augmented large language models: a
review. Mathematics.
10. F Neha, D Bhati & DK Shukla (2025) Retrieval-augmented generation (RAG) in healthcare: A
comprehensive review. AI.
11. N Wiratunga, R Abeyratne, L Jayawardena, et al. (2024) CBR-RAG: case-based reasoning for retrieval
augmented generation in LLMs for legal question answering. International Conference on Case-Based
Reasoning.
12. A Alsubayhay & M Abdalla (2024) Enhancing citizen engagement in E-government services through
AI-driven chatbots. Sebha University Conference Proceedings.
13. L Wang, M Tan & J Han (2016) FastHybrid: A hybrid model for efficient answer selection. Proceedings
of COLING.
14. G Hui & MR Hayllar (2010) Creating public value in e-Government: A public-private-citizen
collaboration framework in Web 2.0. Australian Journal of Public Administration.
15. V Leelavathi & S Pavithra (2026) AI-Based Prediction of Beneficiary Eligibility for Government
Welfare Schemes. International Journal of Engineering and Technical Research.
16. L Breiman (2001) Random forests. Machine Learning.
17. X Wu, Y Gao & D Jiao (2019) Multi-label classification based on random forest algorithm for non-
intrusive load monitoring system. Processes.
18. H Touvron, L Martin, K Stone, P Albert, et al. (2023) Llama 2: Open foundation and fine-tuned chat
models. arXiv preprint arXiv:2307.09288.
19. M Douze, A Guzhva, C Deng, J Johnson, et al. (2025) The FAISS library. IEEE Transactions on Big
Data.
20. V Unnikrishnan & KS Imai (2020) Does the old-age pension scheme improve household welfare?
Evidence from India. World Development.
21. International Institute for Population Sciences (IIPS), et al. (2021) National Family Health Survey
(NFHS-5), India, 2019–21. International Institute for Population Sciences.
22. A Sengupta (2023) Pandemic on employment and earning in urban India during the first three months of
pandemic period: An analysis with unit-level data of periodic labour force survey. The Indian Journal of
Labour Economics.
23. L von Puttkamer (2016) India: Slum-free by 2022? A people-centered evaluation of the Pradhan Mantri
Awas Yojana Scheme. ETH Zurich.
Page 2466
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
24. J Read & F Perez-Cruz (2014) Deep learning for multi-label classification. arXiv preprint
arXiv:1502.05988.
25. T Zhang, V Kishore, F Wu, KQ Weinberger, et al. (2019) BERTScore: Evaluating text generation with
BERT. arXiv preprint arXiv:1904.09675.
26. F Pedregosa, G Varoquaux, A Gramfort, et al. (2011) Scikit-learn: Machine learning in Python. Journal
of Machine Learning Research.
27. M Veale, M Van Kleek & R Binns (2018) Fairness and accountability design needs for algorithmic
support in high-stakes public sector decision-making.
28. C Dwork, M Hardt, T Pitassi, O Reingold, et al. (2012) Fairness through awareness. Proceedings of the
3rd Innovations in Theoretical Computer Science Conference.
29. L Loezer, F Enembreck, JP Barddal, et al. (2020) Cost-sensitive learning for imbalanced data streams.
Proceedings of the 35th Annual ACM Symposium on Applied Computing.
30. G Hovakimyan & JM Bravo (2024) Evolving strategies in machine learning: a systematic review of
concept drift detection. Information.
31. HM Gomes, A Bifet, J Read, JP Barddal, F Enembreck, et al. (2017) Adaptive random forests for
evolving data stream classification. Machine Learning.
32. MK Chan (2013) A dynamic model of welfare reform. Econometrica.
33. N Mehdiyev, C Houy, O Gutermuth, L Mayer, et al. (2021) Explainable artificial intelligence (XAI)
supporting public administration processes – on the potential of XAI in tax audit processes. International
Conference on Electronic Government.
34. SM Lundberg & SI Lee (2017) A unified approach to interpreting model predictions. Advances in Neural
Information Processing Systems.
35. S Kumar, S Datta, V Singh, D Datta, SK Singh, et al. (2024) Applications, challenges, and future
directions of human-in-the-loop learning. IEEE Access.
