A LLM-Powered Agent for Summarizing Critical Information in the

Swine Certiﬁcation Process

Gabriel Rodrigues da Silva

, Alencar Machado

and Vinícius Maran

Laboratory of Ubiquitous, Mobile and Applied Computing (LUMAC), Federal University of Santa Maria,

Santa Maria, Brazil

Keywords:

LLM-Powered Agent, Conversational Assistant, Animal Farm Certiﬁcation.

Abstract:

Animal production farms play an essential role in sanitary control by serving as the ﬁrst line of defense

against disease outbreaks, thus safeguarding both the national food supply and public health. In the state of

Rio Grande do Sul, one of Brazil’s leading regions for livestock production, swine farming holds particular

economic importance and requires rigorous oversight to maintain herd health and compliance with regulatory

standards. Recognizing this critical need, this paper presents the development of a virtual assistant aimed at

supporting swine certiﬁcation processes within the Animal Health Defense Platform of Rio Grande do Sul

(PDSA-RS), a system integral to monitoring and preserving swine health in the region. The virtual assistant

was implemented in Java using Spring Boot and the Spring AI library, with large language models (LLMs)

executed locally through Ollama to ensure data privacy and provide contextualized responses. To improve

response accuracy and relevance, retrieval-augmented generation (RAG) was employed, enriching user queries

with external data on swine health regulations, standard operating procedures, and relevant certiﬁcations.

A case study was conducted to evaluate the effectiveness of the prototype in real-world swine certiﬁcation

scenarios. Results indicated that the virtual assistant showed promise in improving the speed and accuracy

of the certiﬁcation process, offering timely and relevant information to users. This highlights the system’s

potential to streamline workﬂows and facilitate better decision-making among technicians and veterinarians

involved in sanitary control measures.

1 INTRODUCTION

Swine certiﬁcation in Brazil is pivotal to maintain-

ing and expanding the nation’s standing as one of

the world’s top pork exporters. By ensuring com-

pliance with strict global standards for quality, food

safety, and animal welfare, certiﬁcation fosters trust

among trading partners and consumers, signiﬁcantly

contributing to the growth of the swine industry. In

2022 alone, Brazil’s gross production value for swine

reached about R$ 34.175 billion, with the state of Rio

Grande do Sul accounting for nearly 20% of national

slaughter (ABPA, 2024).

In recent years, artiﬁcial intelligence (AI) has ac-

celerated rapidly, fueled by advancements in data

processing and increasingly sophisticated algorithms.

Large-scale language models, such as EcoAssistant,

have gained prominence for performing complex nat-

https://orcid.org/0009-0000-2705-2810

https://orcid.org/0000-0003-2462-7353

https://orcid.org/0000-0003-1916-8893

ural language processing tasks more efﬁciently and

accurately, while reducing operational costs (Zhang

et al., 2023). Various architectures—autoregressive

models (GPT-3), encoder-decoder (BERT), and mul-

timodal models (CLIP)—demonstrate the remarkable

versatility of LLMs in applications beyond natural

language processing (Raffel et al., 2020; Radford

et al., 2021). Notably, GPT-3.5 has shown state-

of-the-art performance in generating coherent, high-

quality text from large datasets (Brown et al., 2020),

while LLaMA caters to contexts where computational

resources are constrained (Touvron et al., 2023).

LLMs have already been adopted across many do-

mains, from virtual assistants to livestock manage-

ment. Ávila (2022) discusses how LLMs and related

AI tools can optimize herd management, reduce ani-

mal stress, and enhance efﬁciency (Shi et al., 2021).

In customer service, LLM-powered virtual assistants

demonstrate more natural and personalized interac-

tions than traditional chatbots, which handle only ba-

sic queries (Peters et al., 2018). Nonetheless, chal-

Rodrigues da Silva, G., Machado, A. and Maran, V.

A LLM-Powered Agent for Summarizing Critical Information in the Swine Certiﬁcation Process.

DOI: 10.5220/0013473400003929

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 965-972

ISBN: 978-989-758-749-8; ISSN: 2184-4992

965

lenges remain in addressing diverse cultural and lin-

guistic contexts and ensuring users fully understand

model outputs (Brown et al., 2020; Touvron et al.,

2023). Further development is critical, particularly

for precision livestock farming and related industries.

This paper is structured as follows: Section 2 cov-

ers the fundamental concepts behind our research.

Section 3 presents the related work, examining pre-

vious studies and solutions relevant to this domain.

In Section 4, we introduce our LLM-powered agent

for the swine certiﬁcation process. Section 5 focuses

on the evaluation of our approach, and Section 6 con-

cludes the paper.

2 BACKGROUND

2.1 Domain-Speciﬁc Agents

Domain-speciﬁc agents are specialized software sys-

tems designed to leverage focused knowledge bases

and processes relevant to a particular ﬁeld. In vet-

erinary certiﬁcation, for instance, these agents must

manage and interpret legislation, health records, im-

munization data, and inspection guidelines with pre-

cision (Peters et al., 2018). Unlike general chat-

bots, which often rely on generic language models or

rule-based ﬂows, domain-speciﬁc agents—especially

those powered by LLMs—deliver more nuanced dia-

logues and can adapt to rapidly changing regulations.

These agents may integrate external data sources,

such as ofﬁcial veterinary databases, to implement ro-

bust Retrieval-Augmented Generation (RAG) strate-

gies. This connection ensures that the agent’s answers

remain accurate and authoritative, reducing the poten-

tial for misinformation. However, creating and main-

taining high-quality, domain-relevant training data

poses signiﬁcant challenges, requiring collaboration

between domain experts and AI developers. Despite

these hurdles, ongoing advances in natural language

processing and knowledge representation promise to

further enhance the reliability and impact of domain-

speciﬁc agents in veterinary public health and animal

certiﬁcation

2.2 LLM

Large Language Models represent a major break-

through in AI, particularly in natural language pro-

cessing. Their development is part of a long-standing

evolution in the ﬁeld, where researchers have endeav-

ored to create machines capable of understanding and

generating human language automatically. Early ef-

forts with neural networks and transformers, such as

GPT (Generative Pre-trained Transformer), paved the

way for GPT-2. Released in 2019 by OpenAI, GPT-2

gained attention by producing coherent and persua-

sive text from simple prompts (Brown et al., 2020).

Currently, LLMs permeate everyday applications,

such as Alexa, Siri, and Google Assistant, enabling

more natural language interactions (Radford et al.,

2019). They also underpin large-scale operations,

from automated customer support and content gener-

ation to medical data analysis (Raffel et al., 2020). A

key factor behind their capabilities lies in billions of

adjustable parameters, which allow models like GPT-

3 to capture linguistic nuances. However, increasing

the number of parameters also raises computational

costs, both during training and inference, and requires

extensive datasets to avoid overﬁtting (Brown et al.,

2020).

Training methods vary. A typical approach is

to ﬁrst pre-train the model on vast amounts of gen-

eral data, imparting broad language skills. Fine-

tuning then adapts these skills to domain-speciﬁc

tasks, such as translation or customer service (Brown

et al., 2020), while in-context learning further reﬁnes

models based on a few specialized examples. Fig-

ure 1 illustrates how an LLM can shift from generic

language tasks to targeted applications.

Figure 1: Example of a model’s adaptability (Bommasani

et al., 2021).

Nonetheless, these architectures bring certain

challenges, especially regarding computational re-

sources and data privacy. Balancing model complex-

ity with efﬁciency is therefore a central concern for

applying LLMs in real-world solutions.

2.2.1 Fine-Tuning

Unlike pre-training with large, general-purpose

datasets, ﬁne-tuning adjusts the model using a

smaller, domain-speciﬁc set of labeled data. This

technique is particularly useful for tasks requiring

specialized language understanding (e.g., sentiment

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

966

analysis or legal text generation), as it optimizes

weights for domain-speciﬁc nuances (Howard and

Ruder, 2018). In a simpliﬁed workﬂow, a pre-trained

model undergoes additional training on a specialized

dataset, leveraging prior knowledge while reducing

overall computational cost (Devlin, 2018).

However, ﬁne-tuning also presents challenges. If

the labeled dataset is insufﬁcient or biased, the model

may inherit these ﬂaws, resulting in skewed or overly

specialized responses. Additionally, ﬁne-tuned mod-

els risk over-ﬁtting, losing the ability to generalize to

new data (Raffel et al., 2020). Maintaining a careful

balance between specialization and broader applica-

bility is therefore essential for successful ﬁne-tuning.

2.2.2 Prompt Engineering

Creating and structuring the instructions provided

to LLMs is an essential practice for optimizing the

responses these systems generate, ensuring that an

LLM’s outputs align with the speciﬁc needs of the

user. With the growing inﬂuence of LLMs, such

as GPT-3, small variations in instruction design

can signiﬁcantly affect model performance (Brown

et al., 2020), underscoring the importance of carefully

crafted prompts.

Beyond guiding the model toward desired behav-

iors, prompt engineering can also adapt the model to

different contexts and audiences without additional

training. This ﬂexibility is particularly useful when

computational resources or domain-speciﬁc data are

limited. By reﬁning prompts, LLMs can perform

complex tasks more efﬁciently and in a more person-

alized manner, reinforcing the value of prompt engi-

neering in real-world AI solutions (Reynolds and Mc-

Donell, 2021).

2.2.3 Retrieval-Augmented Generation (RAG)

Retrieval-Augmented Generation (RAG) signiﬁcantly

advances the integration of language models with in-

formation retrieval, enhancing both accuracy and rel-

evance in generated answers (Lewis et al., 2020). By

combining text generation with external data, RAG

expands the context and applicability of LLMs be-

yond their internal knowledge. As shown in Figure 2,

a prompt and query ﬁrst retrieve relevant information

from external sources, such as legislative or special-

ized databases, which is then integrated into the initial

context before the LLM generates a ﬁnal textual an-

swer. This process allows the model to align text gen-

eration with up-to-date data, overcoming limitations

of pre-trained-only approaches.

RAG typically involves two collaborative mod-

ules: (1) an information retrieval system, essential

Figure 2: How RAG works (Gupta et al., 2024).

for expanding the model’s context and improving re-

sponse accuracy, and (2) a text generation model

that incorporates retrieved data to produce contex-

tualized answers (Brown et al., 2020). This design

adapts quickly to regulatory changes without frequent

re-training, reducing computational costs (Touvron

et al., 2023; Raffel et al., 2020).

Despite its advantages, RAG depends heavily

on the organization and accuracy of external data

sources—poorly maintained or outdated databases

can undermine answer quality (Lewis et al., 2020).

Response time may also increase when large volumes

of data or high query loads are involved (Zhang et al.,

2023). Lastly, security measures become critical in

sensitive domains, such as sanitary certiﬁcation, to

protect both user data and generated outputs (Brown

et al., 2020).

2.2.4 Embedding Process

Embedding transforms text into dense vectors, cap-

turing semantic and contextual relationships (Reimers

and Gurevych, 2019). In the swine certiﬁcation con-

text, embedding helps retrieve legislative paragraphs

closely aligned with user queries.

2.3 Animal Farm Certiﬁcation

Ensuring swine health in Brazil is vital for global mar-

ket competitiveness. Certiﬁcation processes require

detailed tracking of vaccinations, compliance with

biosafety norms, and continuous monitoring (Brown

et al., 2020). As Brazil is a major pork exporter, these

certiﬁcations uphold both national regulations and in-

ternational standards (Descovi et al., 2021).

2.3.1 The PDSA-RS Platform

The PDSA-RS platform streamlines poultry and

swine certiﬁcation (Descovi et al., 2021). It originally

supported four main portals:

Previously, the PDSA-RS offered only four of its

ﬁve main portals, as shown in Figure 3:

• RT Portal. Aimed at veterinarians accredited by

the SVO, enabling them to perform and represent

the legal activities of farms;

A LLM-Powered Agent for Summarizing Critical Information in the Swine Certiﬁcation Process

967

• Laboratory Portal. Equipped with explicit func-

tionalities for sample reception routines, process-

ing, registration, and the issuance of authenticated

reports by veterinarians and staff from private, ac-

credited, and federal laboratories;

• SVE Portal. Geared toward agents of the Sec-

retariat of Agriculture, Livestock, and Irrigation

of Rio Grande do Sul, handling farm registration,

poultry and swine genetic certiﬁcation, warehouse

management, and other related tasks;

• MAPA Portal. Designed for ﬁnal certiﬁcation ac-

tivities involving all parties in certiﬁcate issuance,

result analysis, farm history tracking, and other

responsibilities under Brazil’s Ministry of Agri-

culture, Livestock, and Food Supply.

As new demands arose and additional projects

were implemented within PDSA-RS, the platform ex-

panded its known portals by adding the new Citizen

Portal. This portal is intended for use by any individ-

ual, with streamlined access through gov.br authenti-

cation.

Figure 3: Structure of the PDSA-RS platform.

3 RELATED WORK

The application of AI in livestock management has

seen increasing research interest, particularly in pre-

dictive analytics and automated monitoring systems.

Ávila (2022) explored neural networks for herd op-

timization, demonstrating improvements in health

monitoring and resource allocation (Shi et al., 2021).

Similarly, Mallinger et al. (2024) examined how Ex-

plainable AI (XAI) can facilitate the adoption of smart

farming technologies by improving requirement anal-

ysis and technology design, ensuring that AI-driven

solutions are transparent and user-friendly (Mallinger

et al., 2024). However, most studies focus on struc-

tured data analysis rather than language-based AI

models.

LLMs have been extensively used in customer ser-

vice (Peters et al., 2018) and healthcare applications

(Esteva et al., 2019), yet their role in regulatory com-

pliance for animal certiﬁcation remains underdevel-

oped. Research on AI-driven automation in legal

and administrative ﬁelds has demonstrated the po-

tential of LLMs in document processing and compli-

ance validation (Zhong et al., 2020). The integration

of RAG has further improved accuracy in generat-

ing responses based on external regulatory databases

(Lewis et al., 2020).

Despite advancements in AI-driven certiﬁcation,

existing solutions in livestock management remain

largely manual or rule-based, lacking the contextual

adaptability that LLMs can provide. Our work aims

to bridge this gap by leveraging domain-speciﬁc AI

models to enhance the swine certiﬁcation process

through real-time legal interpretation and compliance

validation.

4 A LLM-POWERED AGENT FOR

SWINE CERTIFICATION

PROCESS

This section describes how we developed and in-

tegrated an LLM-based agent within the PDSA-RS

to aid ofﬁcial veterinaries in swine certiﬁcation pro-

cesses. We detail the software architecture, agent def-

inition, and the steps involved in ensuring contextu-

ally relevant responses.

4.1 Software Architecture Deﬁnition

The architecture of the swine certiﬁcation assis-

tant integrates seamlessly with the PDSA-RS plat-

form, leveraging existing client portals and backend

APIs while introducing new components to support

the LLM-powered assistant. The system uses Java

(Spring Boot) for a stable REST backend, Spring

AI for connecting to LLM services, and Ollama

for the local execution of open-source models (e.g.,

LLaMA). By running LLMs locally, the system en-

sures data privacy—critical for sensitive veterinary

records—while reducing reliance on external APIs.

The architecture is illustrated in Figure 4, which

highlights the components of the PDSA-RS plat-

form. The existing client portals include MAPA,

SVE, RT, and LAB, each serving speciﬁc roles in

swine certiﬁcation and related processes. Among

these, the MAPA portal is the ﬁrst to integrate the

LLM-powered assistant, as indicated by the brain icon

in the image. This marks it as the entry point for test-

ing and deploying the assistant.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

968

Figure 4: Software architecture of the PDSA-RS platform

with LLM integration.

On the backend, the system includes several APIs

for business and service functionalities. Notably, a

new API, the LLM API, was introduced in the ser-

vices layer to facilitate communication with the LLM

model. This API is responsible for handling queries

sent to the model, managing RAG operations, and

retrieving relevant legislative content from the Post-

greSQL database. The database stores vector em-

beddings of regulatory documents to enable semantic

search, as well as logs of user interactions for analyt-

ics and future improvements.

• Backend (Spring Boot). Exposes endpoints to

handle certiﬁcation queries, fetch legislative con-

tent, and manage user sessions.

• Spring AI. Orchestrates prompt engineering and

RAG retrieval. Minimal ofﬁcial documentation

posed integration challenges.

• Ollama. Executes LLMs like LLaMA 3.2:3B

locally, ensuring rapid responses for domain-

speciﬁc queries (OLLAMA, 2024).

This modular design not only ensures the scalabil-

ity and adaptability of the platform but also positions

it to efﬁciently incorporate advanced AI features into

existing workﬂows. The integration with the MAPA

portal demonstrates the potential of the assistant to

enhance certiﬁcation processes while preserving data

privacy and system autonomy.

4.2 Agent Deﬁnition

The agent is a domain-speciﬁc virtual assistant de-

signed to streamline the swine certiﬁcation process by

providing quick and accurate responses to complex

veterinary and regulatory queries. Its primary goal

is to assist veterinarians, inspectors, and other stake-

holders in accessing legislative and procedural infor-

mation necessary for sanitary compliance, thereby

reducing the time and effort required for manual

searches.

The agent operates in a structured manner:

1. Receive a Query. Users, such as veterinarians

or ofﬁcials, input questions related to sanitary

regulations, certiﬁcation requirements, or other

domain-speciﬁc topics.

2. Retrieve Relevant Data (RAG). The system em-

ploys Retrieval-Augmented Generation (RAG) to

search an embedded database containing regula-

tory documents and legislative data. This ensures

that the query is matched to semantically relevant

content.

3. Integrate and Generate a Response. Retrieved

data is appended to the user’s query and passed

to the Large Language Model (LLM). The LLM

synthesizes this information to generate a context-

rich, domain-speciﬁc response.

4. Return Result. The generated response is pre-

sented to the user in an intuitive and accessible

format, providing clear guidance tailored to the

speciﬁc query.

The agent was designed with the unique require-

ments of swine certiﬁcation in mind, ensuring that it

can handle technical terminology, navigate complex

hierarchical regulations, and adapt to evolving leg-

islative standards. However, a notable limitation is

the lack of conversation continuity. Due to the ab-

sence of native multi-turn memory in Spring AI, the

agent processes each query independently. This can

lead to disjointed or repetitive responses in multi-step

interactions, underscoring the need for improved con-

text management mechanisms. Future iterations may

address this limitation by integrating libraries like

Langchain4j or exploring alternative architectures to

support dynamic and continuous dialogue.

By leveraging a combination of advanced in-

formation retrieval and LLM capabilities, the agent

serves as a valuable tool for enhancing decision-

making and operational efﬁciency in swine certiﬁca-

tion workﬂows.

4.3 Prototype Deﬁnition

The prototype for the swine certiﬁcation assistant was

developed to integrate seamlessly with the PDSA-

RS platform, leveraging advanced technologies to

provide efﬁcient and privacy-preserving responses to

user queries. Currently at the proof-of-concept (PoC)

stage, the system aims to validate the feasibility and

potential impact of a locally executed LLM for swine

certiﬁcation.

The software architecture combines a robust back-

end, a curated knowledge base, and a user-friendly in-

terface to streamline the certiﬁcation process for vet-

erinarians and ofﬁcials. The backend, implemented

A LLM-Powered Agent for Summarizing Critical Information in the Swine Certiﬁcation Process

969

using Spring Boot, manages user interactions and pro-

cesses queries through a locally deployed LLM us-

ing Ollama. A PostgreSQL database handles seman-

tic search and logs interactions, enabling RAG for

accurate, context-relevant responses. By embedding

knowledge from domain-speciﬁc documents, the sys-

tem ensures quick access to vital information for cer-

tiﬁcation procedures.

The user interface was extended to integrate a chat

component, as shown in Figure 5. This component

allows users to submit queries, receive real-time an-

swers, and interact with the assistant intuitively. By

embedding this functionality directly into the PDSA-

RS platform, the assistant becomes a practical tool

for stakeholders, simplifying the certiﬁcation work-

ﬂow and improving decision-making efﬁciency.

Figure 5: Chat component integrated into the PDSA-RS

platform. Source: Author’s own work.

The system’s backend architecture, illustrated in

Figure 6, employs the Reactor project’s Flux class to

enable real-time data exchange through ServerSen-

tEvent (SSE). This design allows responses to be

streamed incrementally, enhancing the user experi-

ence by reducing latency and enabling seamless in-

teraction with the assistant.

This prototype demonstrates the feasibility of

deploying LLM-based solutions for domain-speciﬁc

challenges, such as swine certiﬁcation. By combin-

ing advanced AI with tailored infrastructure, the sys-

tem provides veterinarians and technicians with an ef-

fective, user-friendly tool to navigate complex certiﬁ-

cation requirements. The next section illustrates the

assistant’s practical applications through a real-world

usage scenario, highlighting its beneﬁts and identify-

ing potential areas for future improvement.

Figure 6: REST Controller architecture for managing chat

interactions.

5 EVALUATION AND

DISCUSSION

This section presents a comprehensive evaluation of

the prototype, detailing the implementation, a real-

world usage scenario, and a discussion of the results.

5.1 Usage Scenario

Case: Export Inconsistencies. Suppose a veterinar-

ian named João at MAPA faces an urgent export re-

quest for pigs whose vaccination records are incom-

plete. Unsure which vaccines must be immediately

administered according to Brazil’s swine certiﬁcation

rules, João queries the integrated assistant:

“What vaccinations are mandatory for breed-

ing swine in Brazil before export?”

The system:

1. Retrieves relevant normative instructions from the

embedded database.

2. Merges the retrieved content with João’s prompt.

3. LLaMA 3.2:3B generates a response summariz-

ing the mandatory vaccines and references ofﬁcial

documents.

4. Answer: "In Brazil, vaccination is mandatory for

pig farming before export. The speciﬁc vaccines

required may vary depending on the type of pig

and the destination of the product...".

If João wants to ask a follow-up, like scheduling

details or exceptions for certain export destinations,

the assistant may struggle because each new request

is processed independently. This limitation often re-

sults in partial or repetitive answers. Despite that, it

still saves João time, offering specialized guidance far

more efﬁciently than manually searching legislation.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

970

5.2 Evaluation and Discussion

The evaluation of the swine certiﬁcation assistant fo-

cused on assessing its performance, usability, and

limitations in real-world scenarios. The prototype

was deployed on a conventional computer equipped

with an AMD Ryzen 5 5600 processor, 48 GB of

RAM, and an NVIDIA 2060 Super graphics card with

8 GB of VRAM. This setup demonstrated the sys-

tem’s capability to achieve efﬁcient local inference

without requiring high-end infrastructure, reinforcing

the feasibility of deploying such solutions in resource-

constrained environments. During testing, the system

was subjected to various queries, including straight-

forward and complex examples, to evaluate its accu-

racy and relevance. For instance, a user might ask:

• "What vaccinations are mandatory for breeding

swine in Brazil before export?": The system re-

trieved relevant regulations from the database,

synthesized the information, and provided a de-

tailed response citing speciﬁc normative instruc-

tions.

In a follow-up query, such as:

• "Are there exceptions to these requirements for ex-

ports to speciﬁc countries?": The assistant strug-

gled with maintaining context due to the lack of

native multi-turn memory, resulting in responses

that required rephrasing the initial query for clar-

ity. This limitation highlights the need for ad-

vanced context management in future iterations.

The assistant was evaluated on several aspects,

summarized below:

Performance. Running LLaMA 3.2:3B locally via

Ollama delivered near real-time responses, thanks to

the GPU acceleration provided by the NVIDIA 2060

Super. The model’s size was optimized for efﬁcient

inference on this setup, avoiding the need for large-

scale infrastructure while maintaining acceptable re-

sponse times and accuracy

Documentation Limitations. The nascent status

of Spring AI posed challenges during development.

Limited documentation required signiﬁcant effort to

discover workable conﬁgurations and hindered the

customization of prompts to sustain conversational

continuity. This constraint underscored the impor-

tance of community-supported libraries and tools for

future projects.

Context Retention. A notable limitation was the sys-

tem’s inability to retain conversational context across

multiple queries. Each query was processed indepen-

dently, which led to disjointed interactions in scenar-

ios requiring multi-step dialogues. While sufﬁcient

for single-turn queries, the lack of memory negatively

impacted the user experience in more complex use

cases.

Advantages. The prototype demonstrated several key

beneﬁts: - Data Security: All sensitive information,

such as farm records and lab reports, remained on lo-

cal servers, ensuring compliance with privacy stan-

dards. - Faster Development: Leveraging Java and

Spring Boot simpliﬁed integration with existing enter-

prise systems, reducing the time required for develop-

ment. - Scalability: The containerized approach with

Docker and Docker Compose facilitated horizontal

scaling, making the system adaptable to increased us-

age demands.

Overall. The results validated the prototype’s poten-

tial to reduce the effort required to access regulatory

information for swine certiﬁcation. Despite limita-

tions in conversational memory and documentation,

the assistant provided accurate and timely responses

for stand-alone queries. Future improvements should

prioritize enhancing conversational capabilities and

integrating more robust documentation to support de-

velopers and end-users alike

6 CONCLUSION

The integration of LLMs into speciﬁc domains, such

as swine certiﬁcation, addresses the pressing need for

tools that can process and interpret complex, domain-

speciﬁc information efﬁciently. This project pro-

posed and developed a swine certiﬁcation assistant for

PDSA-RS using Java, Spring Boot, Spring AI, and

Ollama. The objective was to demonstrate how lo-

cally deployed LLMs can provide privacy-preserving,

contextualized, and near real-time responses, cater-

ing to the unique requirements of swine health cer-

tiﬁcation in Rio Grande do Sul. The proposal was

motivated by the challenges faced in navigating intri-

cate veterinary regulations and processes, which de-

mand precise, timely decisions. By leveraging RAG

and embedding key regulatory documents, the system

enabled rapid access to relevant information. A case

study validated the prototype’s effectiveness in assist-

ing veterinarians with certiﬁcation tasks, showcasing

its potential to improve decision-making and stream-

line workﬂows.

Despite these advancements, the project encoun-

tered limitations, particularly in maintaining conver-

sational continuity due to the lack of native support

for multi-turn dialogues in Spring AI. Additionally,

the limited documentation for Spring AI integration

posed challenges during development. These barri-

ers highlight opportunities for future improvements,

such as adopting context-tracking frameworks like

A LLM-Powered Agent for Summarizing Critical Information in the Swine Certiﬁcation Process

971

Langchain4j, exploring Python-based ecosystems for

broader tool support, and optimizing model perfor-

mance through compression techniques. In conclu-

sion, this work demonstrated the feasibility and value

of integrating LLMs into the swine certiﬁcation pro-

cess, offering a practical solution for a highly special-

ized domain. Future enhancements to the assistant are

expected to reﬁne its capabilities further, empowering

veterinarians and technicians with accurate, efﬁcient,

and user-friendly tools for managing certiﬁcation pro-

cedures. This study underscores the broader poten-

tial of LLMs to revolutionize domain-speciﬁc appli-

cations, paving the way for innovation in public health

and regulatory compliance.

ACKNOWLEDGEMENTS

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 Inspection

of Animal Origin Products in the State of RS -

UFSM/060496) and FAPERGS, grant n. 24/2551-

0001401-2. The research by Vinícius Maran is par-

tially supported by CNPq, grant 306356/2020-1 -

DT2.

REFERENCES

ABPA (2024). Relatório anual 2024. In Relatório Anual

2024. Associação Brasileira de Proteína Animal.

Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., et al.

(2021). On the opportunities and risks of foundation

models. arXiv preprint arXiv:2108.07258.

Brown, T., Mann, B., Ryder, N., Subbiah, M., et al.

(2020). Language models are few-shot learners. In

Advances in Neural Information Processing Systems,

volume 33, pages 1877–1901. Curran Associates, Inc.

Descovi, G., Maran, V., Ebling, D., and Machado, A.

(2021). Towards a blockchain architecture for animal

sanitary control. In ICEIS (1), pages 305–312.

Devlin, J. (2018). Bert: Pre-training of deep bidirec-

tional transformers for language understanding. arXiv

preprint arXiv:1810.04805.

Esteva, A., Robicquet, A., Ramsundar, B., Kuleshov, V.,

DePristo, M., Chou, K., Cui, C., Corrado, G. S.,

Thrun, S., and Dean, J. (2019). A guide to deep learn-

ing in healthcare. Nature medicine, 25(1):24–29.

Gupta, S., Ranjan, R., and Singh, S. N. (2024). A Com-

prehensive Survey of Retrieval-Augmented Genera-

tion (RAG): Evolution, Current Landscape and Future

Directions.

Howard, J. and Ruder, S. (2018). Universal language model

ﬁne-tuning for text classiﬁcation. arXiv preprint

arXiv:1801.06146.

Lewis, P., Perez, E., Piktus, A., Petroni, F., et al. (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.

Mallinger, K., Corpaci, L., Neubauer, T., Tikász, I. E., et al.

(2024). Breaking the barriers of technology adoption:

Explainable ai for requirement analysis and technol-

ogy design in smart farming. Smart Agricultural Tech-

nology, 9:100658.

OLLAMA (2024). Ollama documentation. Acesso em: 07

dez. 2024.

Peters, F. et al. (2018). Design and implementation of a

chatbot in the context of customer support.

Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., et al.

(2021). Learning transferable visual models from nat-

ural language supervision. In International conference

on machine learning, pages 8748–8763. PMLR.

Radford, A., Wu, J., and Child, R. (2019). Language mod-

els are unsupervised multitask learners. OpenAI blog,

1(8):9.

Raffel, C., Shazeer, N., Roberts, A., Lee, K., et al. (2020).

Exploring the limits of transfer learning with a uniﬁed

text-to-text transformer. Journal of machine learning

research, 21(140):1–67.

Reimers, N. and Gurevych, I. (2019). Sentence-bert: Sen-

tence embeddings using siamese bert-networks.

Reynolds, L. and McDonell, K. (2021). Prompt program-

ming for large language models: Beyond the few-shot

paradigm. In Extended abstracts of the 2021 CHI con-

ference on human factors in computing systems, pages

1–7.

Shi, Y., Wang, X., Borhan, M. S., Young, J., Newman, D.,

Berg, E., and Sun, X. (2021). A review on meat qual-

ity evaluation methods based on non-destructive com-

puter vision and artiﬁcial intelligence technologies.

Food science of animal resources, 41(4):563.

Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux,

M.-A., Lacroix, T., Rozière, B., Goyal, N., Hambro,

E., Azhar, F., et al. (2023). Llama: Open and ef-

ﬁcient foundation language models. arXiv preprint

arXiv:2302.13971.

Zhang, J., Krishnaa, R., Awadallah, A. H., and Wang, C.

(2023). Ecoassistant: Using llm assistant more afford-

ably and accurately. arXiv preprint arXiv:2310.03046.

Zhong, H., Xiao, C., Tu, C., Zhang, T., et al. (2020). How

does NLP beneﬁt legal system: A summary of legal

artiﬁcial intelligence. In Proceedings of the 58th An-

nual Meeting of the Association for Computational

Linguistics, pages 5218–5230, Online. Association

for Computational Linguistics.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

972