Improving Large Language Models Responses with Retrieval Augmented
Generation in Animal Production Certification Platforms
Pedro Bilar Montero
a
, Jonas Bulegon Gassen
b
, Gl
ˆ
enio Descovi
c
, Vin
´
ıcius Maran
d
,
Tais Oltramari Barnasque, Matheus Friedhein Flores
e
and Alencar Machado
f
Laboratory of Ubiquitous, Mobile and Applied Computing (LUMAC), Federal University of Santa Maria, Brazil
Keywords:
Retrieval-Augmented Generation, Large Language Models, Poultry Health, Sanitary Certification, PDSA-RS,
Brazilian Regulations, Legal Texts, Animal Health, Natural Language Processing.
Abstract:
This study explores the potential of integrating Large Language Models (LLMs) with Retrieval-Augmented
Generation (RAG) to enhance the accuracy and relevance of responses in domain-specific tasks, particularly
within the context of animal health regulation. Our proposal solution incorporates a RAG system on the PDSA-
RS platform, leveraging an external knowledge base to integrate localized legal information from Brazilian
legislation into the model’s response generation process. By combining LLMs with an information retrieval
module, we aim to provide accurate, up-to-date responses grounded in relevant legal texts for professionals in
the veterinary health sector.
1 INTRODUCTION
The field of artificial intelligence (AI) has made
significant advancements in recent years, encom-
passing a variety of subfields like computer vision,
robotics, and, most notably, natural language process-
ing (NLP). NLP has especially thrived with the emer-
gence of Large Language Models (LLMs), which
have achieved state-of-the-art results across a range
of tasks, including text generation, summarization,
and language translation. These models, based on
deep neural networks and transformer architectures,
learn from vast corpora and adapt to complex linguis-
tic patterns with minimal supervision, allowing them
to generate coherent and contextually relevant text in
diverse applications (Brown et al., 2020). However,
while LLMs demonstrate broad generalization abili-
ties, they often encounter challenges when applied to
specialized domains, as these areas require extensive,
localized knowledge and precise understanding that is
not always encapsulated within large, generic datasets
a
https://orcid.org/0009-0002-9224-7694
b
https://orcid.org/0000-0001-8384-7132
c
https://orcid.org/0000-0002-0940-9641
d
https://orcid.org/0000-0003-1916-8893
e
https://orcid.org/0000-0003-4436-4327
f
https://orcid.org/0000-0002-6334-0120
(Gururangan et al., 2020; Bommasani et al., 2021)
In industry, the integration of AI into operational
systems has led to significant advancements across
sectors such as healthcare, finance, and legal domains
(Bommasani et al., 2021; Brown et al., 2020). Within
the domain of animal health, for example, LLMs
hold potential to streamline administrative and com-
pliance processes by aiding in decision-making, regu-
latory adherence, and response generation. However,
domain-specific applications introduce complexities
that require accurate, context-aware responses to nu-
anced queries (Gururangan et al., 2020).
The regulatory frameworks governing animal
health, particularly in Brazil, exemplify this chal-
lenge, as professionals must navigate intricate le-
gal and sanitary guidelines to ensure compliance and
protect public health. The Brazilian Minist
´
erio da
Agricultura e Pecu
´
aria (MAPA) plays a key role
in overseeing and enforcing animal health regula-
tions. MAPA’s responsibilities include regulating vet-
erinary practices, approving vaccines, and monitor-
ing the health status of livestock throughout the coun-
try. A specific challenge in animal health regulation in
Brazil is ensuring compliance with the sanitary certi-
fication processes that guarantee the safety of animal
products for both domestic consumption and interna-
tional trade. The certification process, which verifies
that a farm or animal facility meets required sanitary
Montero, P. B., Gassen, J. B., Descovi, G., Maran, V., Barnasque, T. O., Flores, M. F. and Machado, A.
Improving Large Language Models Responses with Retrieval Augmented Generation in Animal Production Certification Platforms.
DOI: 10.5220/0013286100003929
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 1, pages 489-500
ISBN: 978-989-758-749-8; ISSN: 2184-4992
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
489

standards, is particularly critical for industries such
as poultry farming. These certifications ensure that
animal products are free from diseases like avian in-
fluenza, salmonella, and mycoplasmosis, which are
both economically damaging and potentially harmful
to humans.
The Plataforma de Defesa Sanit
´
aria Animal do
Rio Grande do Sul (PDSA-RS) is a platform de-
signed to support the field of animal health regula-
tion in Brazil by implementing an information sys-
tem that integrates all stages of certification processes
for poultry and swine farming in Rio Grande do
Sul. This platform facilitates the organization of pro-
duction activities while ensuring sanitary compliance
with Brazilian animal health regulations. In the case
of the PDSA-RS, veterinary certification processes
depend on the model’s capacity to interpret complex
Brazilian legislation on animal health (Descovi et al.,
2021; Ebling et al., 2024; Schneider et al., 2024).
To address these specialized needs, the Retrieval-
Augmented Generation (RAG) framework has
emerged as a solution for enhancing LLM capa-
bilities by allowing the models to access external
knowledge bases. This framework retrieves relevant
information from a connected knowledge base
during the response-generation process, making it a
valuable approach for domains where specialized,
dynamic knowledge is required. Studies show that
RAG systems can effectively improve the accuracy
of generated responses in specialized fields by
supplementing LLMs with precise, domain-relevant
information (Lewis et al., 2020; Karpukhin et al.,
2020).
In this study, we aim to implement a RAG sys-
tem that can be integrated into the PDSA-RS plat-
form, enhancing its response-generation process with
domain-specific knowledge pertinent to the animal
health regulatory environment in Rio Grande do Sul.
By integrating retrieval mechanisms that draw from a
knowledge base of Brazilian legislation, our system
can produce contextually grounded responses aligned
with the requirements of the animal health regula-
tion in Brazil. This RAG integration holds the po-
tential to bridge the gap between general-purpose
language models and the precise, regulatory-driven
needs of professionals in the veterinary health sector,
contributing to a more effective and reliable AI appli-
cation within the industry.
Our study case will focus on the integration of
this RAG system within the PDSA, specifically in the
module responsible for poultry certification. Our ob-
jective is to create an assistant that will help the pro-
fessional responsible for analyzing poultry certifica-
tion processes to make decisions more quickly. The
purpose of this assistant is to validate all relevant data
for these requests, which will be processed by our
RAG system. Our focus will be to evaluate its perfor-
mance in providing accurate responses to certification
and regulatory questions.
Our paper is organized as follows: Section 2
covers the background of our research, showing all
needed for this paper. In Section 3 we present our
methodology describing how we organized our ar-
chitecture and developed the RAG System. Section
4 is our study case about the implementation of the
RAG System within the PDSA-RS and in Section 5
we show the conclusions of this work and future re-
search possibilities.
2 BACKGROUND
This section provides context for the concepts utilized
in this paper.
2.1 Artificial Intelligence (AI)
Artificial Intelligence (AI) focuses on systems ca-
pable of performing tasks requiring human intelli-
gence. Over decades, AI has progressed remark-
ably, driven by machine learning (ML) and neural
networks, which shifted from symbolic logic-based
methods to data-driven approaches capable of iden-
tifying patterns in unstructured data (Bishop, 2006).
The advent of deep learning enabled significant ad-
vancements, leveraging multilayered neural networks
to process complex data representations effectively
(LeCun et al., 2015). These breakthroughs led to
practical AI applications in healthcare, finance, and
legal compliance (Esteva et al., 2017; Sari and In-
drabudiman, 2024). Among these advancements,
Large Language Models (LLMs) have emerged as
transformative tools for tasks like text generation,
summarization, and translation, powered by innova-
tions like the Transformer architecture (Vaswani et al.,
2017). AI’s potential lies in its adaptability to do-
mains where contextual precision is critical.
2.2 Large Language Models and
Specialized Domains
A Large Language Model (LLM) is a type of artificial
intelligence (AI) designed to process and understand
human language at scale. These models are trained on
vast amounts of text data, enabling them to learn pat-
terns, relationships, and nuances of language. LLMs
have made a lot of progress in recent years, achiev-
ing state-of-the-art results in various NLP tasks such
ICEIS 2025 - 27th International Conference on Enterprise Information Systems
490

as language translation, question answering, and text
generation.
The development of the transformer architecture
by (Vaswani et al., 2017) laid the groundwork for nu-
merous advancements in large-scale language mod-
eling, leading to the creation of influential mod-
els such as GPT by OpenAI. These models demon-
strated that language generation could achieve un-
precedented levels of fluency, coherence, and adapt-
ability in tasks like translation, summarization, and
question answering (Radford and Narasimhan, 2018;
Brown et al., 2020). However, most early LLMs were
proprietary, restricting their accessibility and limiting
the potential for customization and improvement by
the broader research community.
The release of the LLaMA (Large Language
Model Meta AI) family by Meta marked a significant
shift in this paradigm, as it offered high-performance,
large-scale models with an open-access architecture.
LLaMA’s open-source nature allows researchers and
developers to fine-tune and adapt the model for spe-
cific tasks, including specialized domains that require
knowledge and expertise beyond general language
capabilities. This openness has made LLaMA par-
ticularly valuable in academic and applied research
settings, where access to large models with flexible
adaptation options is essential for innovation (Tou-
vron et al., 2023).
However, despite their impressive capabilities,
LLMs often struggle with domain-specific tasks that
require specialized knowledge and context. This lim-
itation arises from several factors:
• Lack of Domain Expertise. While LLMs can be
fine-tuned on specific domains, they may not fully
grasp the intricacies of that domain without exten-
sive training data (Zhang et al., 2023).
• Domain-Specific Nuances. Domains like law,
healthcare, finance, or agriculture involve com-
plex rules, exceptions, and subtleties that may
be difficult for LLMs to capture without explicit
training on these domains (Google Cloud, 2023).
• Limited Contextual Understanding. LLMs rely
on statistical patterns in language to make predic-
tions or generate text. However, this approach can
lead to misinterpretation or misunderstanding of
domain-specific context, leading to inaccurate re-
sults (Brown et al., 2020).
These limitations have led to the development of
hybrid architectures, such as Retrieval-Augmented
Generation (RAG), that aim to enhance LLMs by in-
corporating external knowledge sources to improve
their performance in specific fields (Lewis et al.,
2020; Karpukhin et al., 2020). Unlike fine-tuning,
which requires training the model on a domain-
specific dataset and periodically retraining to update
its knowledge, RAG offers a more dynamic approach
by retrieving relevant information from an external
database or knowledge source in real-time. This ca-
pability is particularly advantageous in fields with fre-
quently updated information, such as legal, medical,
and regulatory domains, where the model’s responses
need to reflect the latest standards and guidelines.
2.3 Retrieval-Augmented Generation
(RAG) Architecture
Retrieval-Augmented Generation (RAG) is an archi-
tecture that enhances large language models (LLMs)
by dynamically combining information retrieval with
text generation, thus allowing LLMs to leverage ex-
ternal knowledge sources while generating responses.
Unlike traditional LLMs that rely solely on pre-
trained knowledge, RAG introduces a retrieval mech-
anism that fetches relevant external documents or data
points, incorporating them into the generation pro-
cess for more contextually accurate responses (Lewis
et al., 2020; Karpukhin et al., 2020).
In RAG, the process starts with transforming user
input into an embedding — a mathematical repre-
sentation of the text. This embedding is then used
to search a vector database, where documents are
pre-processed and stored as embeddings. The vec-
tor database is essential for finding semantically rele-
vant information in response to the user’s query, as
it allows efficient matching of queries with stored
knowledge chunks. Upon retrieval, these documents
are integrated into the generation component of the
model, producing responses that are contextually rel-
evant and informed by up-to-date knowledge sources
(Lewis et al., 2020).
Figure 1 outlines the RAG workflow in four steps:
• 1: The user inputs a query, for example, asking a
question about a recent topic that isn’t present on
the training data of the LLM.
• 2: This step illustrates the indexing of documents
that are split into chunks, encoded into vectors,
and stored in a vector database. The user query
is then used to search this database for relevant
content by a similarity search.
• 3: After the search, the most relevant chunks are
retrieved based on semantic similarity.
• 4: After the search, the most relevant chunks are
retrieved based on semantic similarity.
One of the significant advantages of RAG is its
ability to integrate continuously updated information,
Improving Large Language Models Responses with Retrieval Augmented Generation in Animal Production Certification Platforms
491

Figure 1: Illustration of the RAG Architecture: Information retrieval and generation process. Adapted from (Gao et al., 2023).
making it particularly valuable in fields that require
timely data, such as regulatory environments where
guidelines frequently change. By dynamically ac-
cessing relevant external content, RAG reduces the
risk of outdated or inaccurate responses, improving
both the accuracy and relevance of the model’s out-
put. Furthermore, RAG minimizes the computational
demands associated with continuous fine-tuning, as
it allows the model to access and integrate domain-
specific knowledge on demand rather than retraining
on static datasets (Izacard and Grave, 2021; Lewis
et al., 2020).
In summary, RAG’s approach of combining re-
trieval and generation enables LLMs to adapt to spe-
cific knowledge requirements, providing more robust
support in specialized domains that benefit from up-
dated and context-sensitive responses.
2.4 Fine-Tuning vs
Retrieval-Augmented Generation
(RAG)
In this section, we delve deeper into whether retrieval
modules, as employed in RAG, provide a more robust
benchmark than fine-tuning techniques for achieving
domain-specific accuracy and efficiency.
2.4.1 Retrieval as a Benchmark for
Domain-Specific Accuracy
As Section 2.3 outlines, RAG serve as a dynamic
component by enabling LLMs to access external, up-
to-date knowledge bases in real time. This capability
contrasts with fine-tuning, which relies on embedding
static domain knowledge into the model. Retrieval’s
real-time adaptability ensures that responses remain
accurate in evolving fields such as legal or regulatory
environments.
2.4.2 Efficiency in Computational Resource
Utilization
Fine-tuning techniques, while effective, demand ex-
tensive computational resources and data curation.
QLoRA, for instance, reduces these requirements
by fine-tuning low-rank adaptation layers (Dettmers
et al., 2024), yet it still requires periodic retraining to
incorporate new domain knowledge. Retrieval, on the
other hand, bypasses this need by separating knowl-
edge storage from the generative process, thereby
minimizing overhead and ensuring efficient use of re-
sources (Lewis et al., 2020).
2.4.3 Comparative Evaluation Metrics
When considered as a tool for achieving domain-
specific accuracy, retrieval modules excel in:
ICEIS 2025 - 27th International Conference on Enterprise Information Systems
492

• Adaptability. Retrieval enables models to re-
spond to domain-specific queries with real-time
context, making it better suited for fields with dy-
namic knowledge requirements.
• Scalability. By offloading knowledge storage to
external databases, retrieval reduces the need for
model scaling, unlike fine-tuning which often re-
quires larger model sizes to capture domain intri-
cacies.
• Benchmarking Potential. Retrieval serves as an
ongoing benchmark by continuously updating its
knowledge base, allowing real-world validation of
LLM performance in specialized domains.
From a theoretical standpoint, retrieval modules
highlight a paradigm shift in LLM optimization by
decoupling knowledge retrieval from generative ca-
pabilities. While fine-tuning such as QLoRA embeds
domain-specific expertise into the model, retrieval
treats knowledge as an external, modular component.
This distinction positions retrieval as not merely a
complement to fine-tuning but as a potential alterna-
tive benchmark for evaluating domain-specific effec-
tiveness.
2.5 PDSA-RS and Animal Health
Regulations in Brazil
Established in 2019, the Plataforma de Defesa
Sanit
´
aria Animal do Rio Grande do Sul (PDSA-RS)
supports animal health and production in Rio Grande
do Sul through a real-time digital platform for manag-
ing certifications and ensuring compliance with sani-
tary regulations. Developed by the Federal University
of Santa Maria and MAPA, with Fundesa’s support, it
enhances biosecurity, traceability, and export facilita-
tion.
The system’s modules, such as the poultry health
certification feature, streamline data collection and
certification issuance, linking veterinary inspections,
laboratories, and the agricultural defense authorities.
This interconnected system helps officials monitor
disease control in flocks and facilitates efficient re-
sponse to health risks. PDSA-RS allows inspectors
and producers to follow up on health tests, sam-
ple processing, and the issuance of certificates re-
quired for both domestic and international movement
of poultry, ensuring that health standards are met con-
sistently.
As illustrated in figure 2, the platform adopts
a microservices-oriented architecture. The front-
end comprises several specialized portals tailored for
different stakeholders: the State Veterinary Service
(SVE), technical managers (RTs), agricultural labo-
ratories, and the Ministry of Agriculture (MAPA).
On the back-end, the architecture differentiates
between two distinct types of REST APIs. The busi-
ness APIs manage the core logic and processes asso-
ciated with the platform’s regulatory functions, ensur-
ing that workflows and data management align with
specific legal and procedural requirements. In con-
trast, the service APIs provide more generic func-
tionality, supporting integration and interoperability
with the business APIs by delivering reusable services
across the platform.
This study introduces a dedicated service API to
integrate the platform with a RAG System, facilitat-
ing retrieval and synthesis of regulatory knowledge,
as detailed in subsequent sections.
3 METHODOLOGY
This section outlines the core components of our RAG
System architecture, including the setup of the RAG
Module and the tool used to run our LLM.
3.1 RAG Module
The RAG Module is a structured system designed to
ingest, store, and retrieve documents content based
on semantic similarity. It primarily functions to im-
prove response accuracy by retrieving relevant doc-
ument segments that align with user queries, utiliz-
ing vector embeddings for efficient similarity-based
search to later be fed into the LLM.
3.1.1 Document Ingestion
The ingestion process prepares documents for storage
in the vector database. This involves multiple steps:
• Text Extraction. Document content is extracted
from various file types (PDFs, text files, etc.),
preparing it for chunking and processing.
• Chunking and Tokenization. Documents are di-
vided into smaller sections, such as sentences or
paragraphs. Chunking improves retrieval preci-
sion and allows targeted access to specific infor-
mation. Tokenization follows, breaking down text
into discrete tokens for embedding..
• Embedding Creation. Each chunk is trans-
formed into a vector embedding using a pre-
trained embedding model, which captures the se-
mantic content of the text. The embedding is a
high-dimensional numerical representation of the
chunk, enabling similarity-based retrieval within
the vector database.
Improving Large Language Models Responses with Retrieval Augmented Generation in Animal Production Certification Platforms
493

Figure 2: Overview of the PDSA-RS Architecture.
We choose to use the LlamaIndex data framework
for this stage, leveraging the Python package llama-
index. This package provides ready-to-use functions
for document ingestion, which simplifies the process
of embedding documents, indexing them, and storing
them in a vector database for efficient retrieval.
3.1.2 Vector Storage
The RAG module utilizes a vector database to store
embeddings. This type of storage allows efficient re-
trieval through similarity search, where embeddings
with high semantic similarity to a query are located
based on distance metrics. PostgreSQL was selected
as the database solution, using the PGVector exten-
sion to store embeddings. PGVector is a PostgreSQL
extension that provides efficient support for vector-
based data storage, making it well-suited for handling
the embeddings generated by the LlamaIndex frame-
work.
3.1.3 Retrieval Process
The RAG module’s retrieval component is used when
a user submits a query. The query is converted into
an embedding and used to locate relevant document
chunks stored in the vector database:
• Query Pre-Processing. Before retrieval, the
query may undergo filtering and rephrasing to
enhance search relevance. Steps may include
expanding acronyms, removing stop words, and
restructuring complex queries into simpler sub-
queries. These adjustments help the RAG mod-
ule better interpret the query’s intent and improve
retrieval accuracy.
• Similarity Search. Using the query embed-
ding, a similarity search is conducted in the vec-
tor database. This search identifies chunks that
closely align with the query’s semantic content.
Cosine similarity or other distance-based metrics
are applied to find the most relevant chunks, en-
suring that only document sections pertinent to the
query are retrieved.
We are again utilizing LlamaIndex for the retrieval
process, taking advantage of its built-in search func-
tions. LlamaIndex provides efficient methods for per-
forming semantic searches over document embed-
dings stored in the vector database. These search
functions allow us to find the most relevant document
chunks based on the similarity between the query em-
bedding and the stored document embeddings.
3.1.4 Maintenance and Updates
To keep information relevant and accurate, the RAG
module may undergo periodic updates:
• Embedding Updates. As new documents be-
come available, they are embedded and stored in
ICEIS 2025 - 27th International Conference on Enterprise Information Systems
494

the vector database to expand the scope of retriev-
able content.
• Maintenance of Vector Store. Regular mainte-
nance helps optimize performance and relevance.
Outdated information can be removed, and fre-
quently accessed chunks may be optimized for
faster retrieval, ensuring that the vector store re-
mains responsive and accurate.
3.2 Large Language Model Setup
The LLM selected for use in our RAG system is the
Llama 3.1 8B Instruct model. This model, part of
Meta’s LLaMA family, is a fine-tuned model from the
base Llama 3.1 8B, designed for instruction-following
tasks. We preferred the 8B version as it has a balance
between computational efficiency and performance,
offering a strong capacity for understanding and gen-
erating responses while also maintaining an efficient
power consumption.
3.2.1 Model Selection and Configuration
To run our LLM, we have chosen to use privateGPT
in conjunction with the Ollama framework. This com-
bination allows us to maintain complete control over
the model and data, ensuring privacy and security.
The privateGPT is a solution designed to run mod-
els locally, ensuring that the data and queries remain
private without needing to send sensitive information
to external servers. It is useful when working with
confidential or proprietary data, like legal or medical
information. In this setup, privateGPT operates as the
interface to manage the LLM, which is hosted on a
private server, offering full control over the model.
To run the model locally, privateGPT offers two
frameworks: llama.cpp and Ollama. We chose Ol-
lama based on the privateGPT documentation, which
recommends it for its greater versatility in running
across different computational environments. This
framework is designed to simplify model deployment,
providing an efficient and flexible setup that can be
easily adapted to various hardware configurations.
3.2.2 API
To allow external services to interact with our system,
we chose to use an API. PrivateGPT already includes
an API as part of its structure, making it easy to in-
tegrate our system with external services. This API
provides all the necessary endpoints for document in-
gestion, query handling, and response generation, ab-
stracting the complexities of the RAG pipeline while
allowing external systems to make requests and re-
ceive responses in a standardized way.
This API is divided into two logical blocks:
• High-Level API. Abstracts the complexities of
the RAG pipeline by automating document pro-
cessing and response generation. It manages
tasks such as document parsing, splitting, meta-
data extraction, embedding generation, and stor-
age, preparing documents for efficient retrieval.
Additionally, the API handles chat and comple-
tion processes by retrieving relevant content, en-
gineering prompts, and generating responses, al-
lowing users to focus on querying the system and
receiving contextually relevant answers based on
the ingested documents, without needing to man-
age the intricate details of retrieval and genera-
tion.
• Low-Level API. Provides advanced users with
the ability to generate embeddings for any given
piece of text. It also includes an endpoint for
contextual chunks retrieval, which, when given
a query, searches the ingested documents and re-
turns the most relevant text chunks.
3.3 Architecture Overview
Our architecture can be summarized as shown in fig-
ure 3. The system is composed of the RAG Module,
which is responsible for the entire process of inges-
tion, storage, and retrieval of context, and the LLM,
which processes user queries. The RAG module han-
dles document parsing, embedding generation, and
storing the relevant context in a vector database for
efficient retrieval. The LLM, once the context is re-
trieved, generates responses based on the processed
queries, ensuring the responses are contextually rele-
vant and informative.
4 CASE STUDY
To evaluate the effectiveness of our RAG system on
the PDSA-RS platform, we conducted a case study to
integrate the RAG system within the platform and as-
sess its performance in handling regulatory questions
related to poultry certification.
In order for an establishment to obtain a sanitary
certificate they must submit samples of birds, poul-
try products (such as eggs), among other materials
for laboratory testing in institutions accredited by the
MAPA. The purpose of these collections is to ensure
that the batches are free from pathogenic agents such
as Salmonella or other relevant infectious agents. For
each age group of birds, there are different rules re-
garding the quantity of materials to be collected, the
Improving Large Language Models Responses with Retrieval Augmented Generation in Animal Production Certification Platforms
495

Figure 3: Overview of the RAG System Architecture.
types that must be collected, and the combinations
of each material. Furthermore, the purpose of pro-
duction for these birds must also be considered, as
it influences the aforementioned parameters. We use
these parameters to define the query that will be sent
to the retrieval module to search for context in our
knowledge base.
For the documents used in the analyses performed
by the RAG system, we are utilizing IN 78/2003
(Minist
´
erio da Agricultura, 2003) and IN 44/2001
(Minist
´
erio da Agricultura, 2001). These documents
outline the technical standards for disease control and
certification of poultry establishments free or con-
trolled from diseases like Salmonella and Mycoplas-
mosis.
4.1 Query Structure
We analyzed an object representing a completed certi-
fication request, which includes data on diseases mon-
itored , sanitary conditions and laboratory testings.
In this study, the focus was verifying whether all the
laboratory testing for monitored diseases met the re-
quired standards.
The data structure from a sample certification ob-
ject includes:
• Nucleus Details. Information about the facility,
such as active status, purpose of production (e.g.,
Hatchery), and other relevant parameters.
• Laboratory Exams. The list of tested diseases,
including Pullorum (SP), Gallinarum (SG), and
others, containing dates of each test and also ma-
terials used for testing.
• Sanitary Conditions. Each disease status (e.g.,
Free or Vaccinated).
Our system was tasked to analyze the object, par-
ticularly the laboratory exams, to ensure that all ele-
ments were compliant with the required standards for
certification.
As the context of our work is within a controlled
environment, we have predictability of which data we
want to feed into our LLM, thus we can use pre-
established queries for the context where the module
will be used. In the example of the health certification
process for poultry establishments, we make choices
according to the rules of each stage of the life cycle
of a poultry establishment.
We did some experimentation and arrived at the
following system prompt, which will be used to
provide background instructions on how the model
should answer. This prompt showed little incidence
of hallucinations in the tests carried out:
”You must act as an expert in poultry farming
and certification. Answer questions accurately about
poultry certification and biosecurity. If unsure, indi-
cate that the question cannot be answered.”
For the condensed prompt we followed the struc-
ture below:
ICEIS 2025 - 27th International Conference on Enterprise Information Systems
496

{{ Retrieved Context }}
Use the content above to respond to the query
below if relevant, or respond to the best of
your ability without it.
{{ Poultry Certification Request Data }}
The condensed prompt is designed aiming to
boost contextual awareness and foster model behavior
that prioritizes accuracy by only generating responses
aligned with the provided information.
As shown in table 1, for the parameters of our
LLM, we set the temperature value to 0.8. With this
setting we were able to balance the model’s ability to
explore different possibilities while avoiding trivial or
overly repetitive answers. Temperature is a hyperpa-
rameter that controls the randomness or softness of
output probabilities in LLMs, allowing for more di-
verse or exploratory language generation rather than
deterministic predictions. Additionally, we employed
an adaptive context window size with a maximum ca-
pacity of 1024 tokens, accommodating a wide range
of input lengths while ensuring efficient memory use.
For output search strategies, we implemented a beam
search with a width of 16, enabling the model to eval-
uate multiple possible answers and select the best out-
put path based on probability scores.
Resource allocation was also a focus, with GPU
memory dynamically set at 8GB to allow efficient
processing on available hardware. CPU utilization
was capped at 80%, with a maximum of 16 threads,
mostly because of our limitation with the test hard-
ware, to ensure that the model could make full use of
our CPU power.
Table 1: LLM Hyperparameter Configuration.
Hyperparameter Value
Temperature 0.8
Context Window Size Adaptive (max 1024 tokens)
Beam Search Width 16
GPU Memory Allocation 8GB (dynamic allocation)
CPU Thread Utilization 80% (max 32 threads)
We encountered significant challenges in running
inferences on our test hardware setup, which con-
sists of a Ryzen 7 5800X3D CPU and a Radeon RX
7700XT 12GB GPU. Specifically, we had to carefully
manage GPU memory allocation to prevent excessive
memory usage, as the RX 7700XT’s 12GB VRAM
is limited compared to more powerful professional-
grade GPUs.
The integration of the RAG System with PDSA-
RS is done within one of the platform’s several mi-
croservices, more specifically the service responsible
for the poultry certification process, called Aves API.
Within this API, an endpoint is responsible for mak-
ing the call to the RAG system for chat completion,
containing all the information that is extracted from
PDSA-RS according to the user’s request. As shown
in figure 4, this call is made through a button found
on the front-end of the platform, within the compo-
nent where certificate requests are analyzed.
4.2 Results
We then conducted tests using certification request
data to evaluate the system’s performance. For eval-
uation metrics, we used response accuracy and gen-
eration time. To measure the accuracy of the RAG
system, we extracted real data from certifications that
were produced by specialists that use the PDSA-RS.
We used the condensed prompt described in section
4.1 to each evaluation step, baseline and RAG.
We take the laboratory exams and create a query
to ask the LLM if, based on these exams, all the re-
quired materials were collected on the correct dates,
according to the purpose of production for each stage.
Based on this analysis, we ask the LLM to determine
the sanitary condition of the farm for the tested dis-
ease and if it would be eligible for certification. We
compare the results with the sanitary conditions that
were made by a specialist to verify if the generation
by the LLM was correct.
We categorized responses into two types:
• Fully Correct Responses (Rc): The model cor-
rectly identified compliance or non-compliance
for all nuclei.
• Partially Correct Responses (Rp): The model
provided partially accurate responses, missing or
misinterpreting some details.
The formula used to calculate accuracy was:
Accuracy =

Rc +(W × Rp)
N

× 100
Where:
• N = Total queries in the benchmark
• W = Weight for partially correct responses (set to
0.5 in our case)
For comparison, we generated responses in two
modes: with RAG, using the IN documents incorpo-
rated into our knowledge base, and without RAG, us-
ing only the base model. This approach allowed us
to assess the impact of the RAG module on response
quality and processing efficiency.
For the baseline model, which does not use the
RAG system, the evaluation metrics were as follows:
• N = 100: Total queries related to certification re-
quests.
Improving Large Language Models Responses with Retrieval Augmented Generation in Animal Production Certification Platforms
497

Figure 4: Screenshot of the poultry certification analysis from the PDSA-RS.
• Rc = 2: Fully aligned responses.
• Rp = 20: Partially correct responses.
Out of 100, there were 78 completely incorrect re-
sponses that were not accounted and therefore do not
appear in the above parameters.
Applying the formula for the baseline:
Accuracy =

2 + (0.5 × 20)
100

× 100 = 12%
For the RAG-enhanced system, the evaluation
metrics were:
• N = 100: Total queries related to certification re-
quests.
• Rc = 40: Fully aligned responses.
• Rp = 30: Partially correct responses.
Out of 100, there were 30 completely incorrect re-
sponses that were not accounted and therefore do not
appear in the above parameters.
Applying the formula for the RAG-enhanced sys-
tem:
Accuracy =

40 + (0.5 × 30)
100

× 100 = 55%
As shown in table 2, baseline generation showed a
low accuracy of 12%, which is a result of the LLM’s
Table 2: Model Performance.
Phase Response Accuracy Generation Time (seconds)
Baseline 12% 5.2
RAG 55% 8.3
limited capacity to contextualize and apply domain-
specific knowledge regarding poultry certification and
internal monitoring standards.
In contrast, the RAG demonstrated a significant
improvement in accuracy. The model’s ability to re-
trieve relevant legal texts allowed it to generate re-
sponses that were more informed and contextually ac-
curate.
The generation time is slightly longer when using
RAG, as content retrieval is required before the re-
sponse generation process can begin. This additional
retrieval step, which involves searching for and load-
ing relevant context from the knowledge base, nat-
urally extends the time needed to generate each re-
sponse. However, this trade-off enhances response
accuracy and relevance by incorporating contextual
information.
The results indicate that the integration of a
RAG system substantially improves response accu-
racy, showing an accuracy increase of 43% when us-
ing the RAG approach.
The findings from this case study suggest that inte-
grating Retrieval-Augmented Generation (RAG) into
a platform like PDSA-RS offers a practical solution
for enhancing the performance of LLMs in domain-
ICEIS 2025 - 27th International Conference on Enterprise Information Systems
498

specific tasks as well as assisting and accelerating the
process of analyzing the required standards necessary
for poultry certification.
It is important to emphasize that the objective of
this case study was not to replace human expertise
in the poultry certification process. Rather, the focus
was to assess how a LLM, integrated with a Retrieval-
Augmented Generation (RAG) system, could assist in
streamlining certain tasks by providing relevant le-
gal information and context. While LLMs can offer
valuable support in processing complex data and re-
trieving domain-specific knowledge, they still present
limitations, such as the potential for generating in-
accurate or incomplete responses. Therefore, human
oversight remains crucial to ensure the reliability and
precision of the certification process, particularly in
areas where nuanced judgment and expert knowledge
are required.
5 CONCLUSION
The integration of a Retrieval-Augmented Generation
(RAG) system into the PDSA-RS platform has proven
to be a valuable advancement in improving the accu-
racy and contextual relevance of large language mod-
els (LLMs) within the specialized domain of animal
health regulation. By leveraging domain-specific re-
trieval to supplement the generative capabilities of
LLMs, our system bridges the gap between general-
purpose language models and the unique, nuanced
needs of regulatory environments.
Our case study highlights the potential for RAG
systems to streamline and enhance the way legal and
regulatory queries are handled, especially in complex
sectors like veterinary health. The performance gains
observed through the incorporation of relevant reg-
ulatory texts into the LLM’s output underscore the
value of domain-adapted retrieval processes in in-
creasing both the precision and usefulness of the gen-
erated responses. This approach demonstrates that
RAG not only improves response quality but also pro-
vides a scalable solution to address domain-specific
challenges, where accuracy and legal compliance are
paramount.
Looking ahead, the continuous refinement of both
the retrieval module and the underlying language
model will be critical in ensuring that the PDSA-RS
platform remains capable of delivering accurate and
timely legal guidance.
From a theoretical perspective, this study empha-
sizes the adaptability of RAG in dynamic environ-
ments and its ability to overcome knowledge obso-
lescence—a limitation inherent to fine-tuned models.
Furthermore, the findings contribute to the broader
discourse on hybrid AI systems, where generative
and retrieval capabilities are combined to enhance
domain-specific applications.
For future work, there is the potential to fine-
tune the model with specific documents, which could
further enhance the relevance and accuracy of con-
tent generation on this subject. This additional fine-
tuning would allow the model to better handle com-
plex, domain-specific queries, delivering more pre-
cise responses tailored to the chosen topic. Addi-
tionally, developing a graphical user interface (GUI)
could significantly improve user interaction by allow-
ing users to directly input various documents and in-
teract with them. This interface would enable users
to upload documents, ask questions, and receive con-
textually relevant responses, making the system more
intuitive and accessible for end-users engaging with
diverse content.
In conclusion, while the RAG-enhanced LLM sys-
tem offers significant benefits, it is essential to main-
tain human oversight, particularly in legal and reg-
ulatory contexts where misinterpretation of guide-
lines could have serious consequences. The role
of the LLM is not to replace human expertise but
to augment decision-making by providing informed,
contextually relevant suggestions. This hybrid ap-
proach—leveraging both cutting-edge AI and human
expertise—represents a promising path forward for
regulatory platforms like PDSA-RS, fostering innova-
tion while ensuring the accuracy and integrity of the
certification processes that underpin Brazil’s animal
health sector.
ACKNOWLEDGEMENT
This research is supported by FUNDESA, project
“Combining Process Mapping and Improvement with
BPM and the Application of Data Analytics in
the Context of Animal Health Defense and Inspec-
tion of Animal-Origin Products in the State of RS”
(UFSM/060496) and by MPA (Minist
´
erio da Pesca e
Aquicultura), project “Use of Artificial Intelligence
in the Systematization of Hygienic-Sanitary Certifi-
cation Processes for Vessels and Accreditation of Le-
gal Origin of Fish” (UFSM/060642). The research by
Vin
´
ıcius Maran is partially supported by CNPq grant
306356/2020-1 DT-2, CNPq PIBIC and PIBIT pro-
gram and FAPERGS PROBIC program.
Improving Large Language Models Responses with Retrieval Augmented Generation in Animal Production Certification Platforms
499

REFERENCES
Bishop, C. (2006). Pattern Recognition and Machine
Learning, volume 16, pages 140–155.
Bommasani, R., Hudson, D., Adeli, E., Altman, R., Arora,
S., Arx, S., Bernstein, M., Bohg, J., Bosselut, A.,
Brunskill, E., Brynjolfsson, E., Buch, S., Card, D.,
Castellon, R., Chatterji, N., Creel, K., Davis, J., Dem-
szky, D., and Liang, P. (2021). On the opportunities
and risks of foundation models.
Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D.,
and Dhariwal, e. a. (2020). Language models are few-
shot learners. In Larochelle, H., Ranzato, M., Hadsell,
R., Balcan, M., and Lin, H., editors, Advances in Neu-
ral Information Processing Systems, volume 33, pages
1877–1901. Curran Associates, Inc. pp. 33-34.
Descovi, G., Maran, V., Ebling, D., and Machado, A.
(2021). Towards a blockchain architecture for animal
sanitary control. In Proceedings of the 23rd Inter-
national Conference on Enterprise Information Sys-
tems - Volume 1: ICEIS,, pages 305–312. INSTICC,
SciTePress.
Dettmers, T., Pagnoni, A., Holtzman, A., and Zettlemoyer,
L. (2024). Qlora: efficient finetuning of quantized
llms. In Proceedings of the 37th International Con-
ference on Neural Information Processing Systems,
NIPS ’23, Red Hook, NY, USA. Curran Associates
Inc.
Ebling, D., Machado, F., Descovi, G., Cardenas, N.,
Machado, G., Maran, V., and Machado, A. (2024). A
distributed processing architecture for disease spread
analysis in the pdsa-rs platform. In Proceedings of the
26th International Conference on Enterprise Informa-
tion Systems - Volume 2: ICEIS, pages 313–320. IN-
STICC, SciTePress.
Esteva, A., Kuprel, B., Novoa, R., Ko, J., Swetter, S., Blau,
H., and Thrun, S. (2017). Dermatologist-level clas-
sification of skin cancer with deep neural networks.
Nature, 542.
Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y.,
Dai, Y., Sun, J., Guo, Q., Wang, M., and Wang, H.
(2023). Retrieval-augmented generation for large lan-
guage models: A survey. ArXiv, abs/2312.10997.
Google Cloud (2023). A three-step design pattern for spe-
cializing llms. Accessed: 2024-09-13.
Gururangan, S., Marasovi
´
c, A., Swayamdipta, S., Lo, K.,
Beltagy, I., Downey, D., and Smith, N. A. (2020).
Don’t stop pretraining: Adapt language models to do-
mains and tasks. In Jurafsky, D., Chai, J., Schluter,
N., and Tetreault, J., editors, Proceedings of the 58th
Annual Meeting of the Association for Computational
Linguistics, pages 8342–8360, Online. Association
for Computational Linguistics.
Izacard, G. and Grave, E. (2021). Leveraging passage re-
trieval with generative models for open domain ques-
tion answering. In Merlo, P., Tiedemann, J., and Tsar-
faty, R., editors, Proceedings of the 16th Conference
of the European Chapter of the Association for Com-
putational Linguistics: Main Volume, pages 874–880,
Online. Association for Computational Linguistics.
Karpukhin, V., Oguz, B., Min, S., Lewis, P., Wu, L.,
Edunov, S., Chen, D., and Yih, W.-t. (2020). Dense
passage retrieval for open-domain question answer-
ing. In Webber, B., Cohn, T., He, Y., and Liu,
Y., editors, Proceedings of the 2020 Conference on
Empirical Methods in Natural Language Processing
(EMNLP), pages 6769–6781, Online. Association for
Computational Linguistics.
LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learn-
ing. Nature, 521:436–44.
Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin,
V., Goyal, N., K
¨
uttler, H., Lewis, M., Yih,
W.-t., Rockt
¨
aschel, T., Riedel, S., and Kiela,
D. (2020). Retrieval-augmented generation for
knowledge-intensive nlp tasks. In Proceedings of the
34th International Conference on Neural Information
Processing Systems, NIPS ’20, Red Hook, NY, USA.
Curran Associates Inc.
Minist
´
erio da Agricultura, Pecu
´
aria e Abastecimento, S. d.
D. A. (2001). Instruc¸
˜
ao normativa nº 44, de 23 de
agosto de 2001. Accessed: 2024-10-23.
Minist
´
erio da Agricultura, Pecu
´
aria e Abastecimento, S. d.
D. A. (2003). Instruc¸
˜
ao normativa nº 78, de 3 de
novembro de 2003. Accessed: 2024-10-23.
Radford, A. and Narasimhan, K. (2018). Improving lan-
guage understanding by generative pre-training.
Sari, Y. and Indrabudiman, A. (2024). The role of artificial
intelligence (ai) in financial risk management. For-
mosa Journal of Sustainable Research, 3:2073–2082.
Schneider, R., Machado, F., Trois, C., Descovi, G., Maran,
V., and Machado, A. (2024). Speeding up the simu-
lation animals diseases spread: A study case on r and
python performance in pdsa-rs platform. In Proceed-
ings of the 26th International Conference on Enter-
prise Information Systems - Volume 2: ICEIS, pages
651–658. INSTICC, SciTePress.
Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux,
M.-A., Lacroix, T., Rozi
`
ere, B., Goyal, N., Hambro,
E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E.,
and Lample, G. (2023). Llama: Open and efficient
foundation language models.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,
L., Gomez, A. N., Kaiser, L. u., and Polosukhin,
I. (2017). Attention is all you need. In Guyon,
I., Luxburg, U. V., Bengio, S., Wallach, H., Fer-
gus, R., Vishwanathan, S., and Garnett, R., editors,
Advances in Neural Information Processing Systems,
volume 30. Curran Associates, Inc.
Zhang, S., Dong, L., Li, X., Zhang, S., Sun, X., Wang, S.,
Li, J., Hu, R., Zhang, T., Wu, F., et al. (2023). In-
struction tuning for large language models: A survey.
arXiv preprint arXiv:2308.10792.
ICEIS 2025 - 27th International Conference on Enterprise Information Systems
500