INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
AI - Driven Internal Data Intelligence Assistant  
Sampada Kulkarni1, Shivi Verma, Apeksha Hatle2, Shreeja Rajput3, Pratiksha Jadhav4  
1,2,3,4Department of Information Technology, Progressive Education Society’s Modern College of  
Engineering, Pune, India  
DOI: https://doi.org/10.51244/IJRSI.2025.1210000288  
Received: 08 November 2025; Accepted: 17 November 2025; Published: 19 November 2025  
ABSTRACT  
The proliferation of generative AI presents significant productivity opportunities for software companies, but  
the practice of employees using public AI chatbots poses severe data security risks. This research paper  
demonstrates the application of a Retrieval-Augmented Generation (RAG) architecture into building a  
dedicated AI assistant for secure, internal corporate use by reasoning its responses exclusively in a company's  
private document repository, unlike a general-purpose model. This will ensure that sensitive internal data such  
as project specifications, internal wikis, project codes, knowledge documents, policy documents etc do not  
leave the company firewall during the AI operation cycle while also letting the company entity use AI to  
compare and access the needful resources from huge internal data.Our methodology involves processing and  
vectorizing internal documents, enabling semantic search for precise information retrieval, and leveraging a  
large language model (LLM) solely for generating context-aware responses from the retrieved data. We argue  
that this system provides a practical, secure, and efficient solution for knowledge management and employee  
assistance, balancing the power of modern AI with the non-negotiable demands of corporate data privacy.  
Keywords: Retrieval-Augmented Generation (RAG), AI data security, Knowledge management, vector  
database, code generation, Natural language Processing (NLP), Document intelligence  
INTRODUCTION  
Today's software development environment is distinctly information-intensive. Development teams, project  
managers, and quality assurance engineers often deal with the challenge of locating and understanding internal  
documentation, the context of complicated legacy codebases which can be poorly documented, and changing  
company policy or security standards. Having to locate, integrate, apply, and translate all of this knowledge is  
a drag on productivity and operational efficiency. Concurrently, large language models (LLMs) have caused a  
disruptive entry point for natural language-based interaction, including the potential to change how humans  
engage with and obtain information. Examples of public-facing artificial intelligence (AI) assistants  
demonstrate impressive capabilities, however the opportunity for corporate environments is constrained by a  
serious security risk, specifically, they require proprietary data— intellectual property, trade secrets or  
sensitive customer data to be sent to external, third-party servers  
Accordingly, establishing this type of an attack vector is unacceptable, it puts organizations at risk of data  
breach, compliance with data protection obligations and puts them at a competitive disadvantage. In summary,  
the promise of AI efficiency is intriguing, but often the levels of risk associated with public tools just does not  
allow for use in enterprise. This article proposes the architecture for a private, self-contained AI assistant to  
combat this significant conflict. The fundamental challenge is to leverage the analytical and interpretive  
capabilities of AI to amplify human capacity for information retrieval and understanding while also providing  
data confidentiality in perpetuity. The solution proposed here removes dependence on third-party AI services  
by implementing a Retrieval-Augmented Generation (RAG) system that runs exclusively on a company's  
private document repository. This solution creates a closed-loop, secure environment where the AI's  
knowledge is intentionally limited to internal vetted/approved information. By interrogating and processing  
information entirely inside the organizational perimeter, the RAG solution provides the benefit of intelligent  
assistance without compromising data sovereignty.The forthcoming sections will outline the limitations of  
other systems, describe our approach to building a local RAG-based assistant, and offer a full architectural  
schema. The paper will conclude with a review of the effectiveness of the system in achieving its bi-focal goal  
Page 3318  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
of providing augmented efficiency with absolute data security.  
Objective  
The primary goal of our project is to design, implement, evaluate, test and assess a completely local AI  
assistant that provides secure and effective document intelligence for organisations. Our project will use a  
Retrieval-Augmented Generation (RAG) architecture to keep sensitive company information bounded by a  
private infrastructure. This will eliminate the need to use a publicly-available AI service with which sharing  
the information can risk leaking sensitive data. To accomplish this overall goal, the specific objectives shown  
below have been developed:  
1. To design and build a securely working Retrieval-Augmented Generation (RAG) pipeline. The project  
will focus on developing a system that can read documents (e.g., PDFs, Word documents, text files in all  
formats), convert the documents into a vectorized format by using technology that operates locally.  
2. To create a user interface for easy interaction. This objective is to create a web interface where users can  
upload a document and ask their question in the chat tab in natural language easily such that the assistant  
can be instructed to complete a task (for e.g. write a code to add two numbers). The interface should  
include a way to present AI-generated answers from its own data citations, so the answers are transparent  
to the users  
3. To ensure that we maintain data sovereignty and privacy through local LLM use. This critical objective  
highlights integrating a locally-hosted large language model (e.g., via Ollama) for the final response  
generation. By processing all data and AI inferences entirely on local hardware, the system will guarantee  
that no sensitive information is transmitted to or processed by third-party APIs.  
4. To evaluate the system's performance based on statistical measures: accuracy, response time, and  
usability. The final objective of our project is to conduct a detailed assessment of the developed system.  
This will involve measuring the accuracy of the answers against a validated set of queries, standardizing  
the response latency for a satisfactory user experience, and gathering user feedback on platform use.  
LITERATURE SURVEY  
In recent years, to develop the AI-Driven Internal Data Intelligence Assistant, we reviewed several studies  
related to privacy-focused large language models (LLMs), retrieval-augmented generation (RAG), and secure  
enterprise data management. These studies offered valuable understanding of how to protect sensitive data,  
enhance information retrieval, and implement AI solutions safely within organizational environments.  
Earlier research focused either on performance or data security, but very few combined both in one complete  
framework. This review helped identify that gap and guided the design of the proposed system.  
1. DB-GPT: Empowering Database Interactions with Private LLMs (Xue et al., 2024) – Introduced a  
privacy-focused LLM using RAG for database queries. While effective for structured data, it required  
heavy computation and was less suitable for unstructured or lightweight local setups.  
2. Confidential Prompting (Zhang et al., 2023) – Proposed using Trusted Execution Environments (TEEs)  
to secure prompts during cloud inference. Though privacy improved, dependence on cloud infrastructure  
raised costs and limited offline use.  
3. E2E Data Extraction Framework (Kumar & Singh, 2025) – Developed a deep learning-based system to  
structure unorganized enterprise data. However, it did not include security or access control features  
needed for sensitive information.  
4. Fine-Tuning LLMs for Enterprise Applications (Raj et al., 2024) – Explored efficient fine-tuning  
methods like LoRA and Q-LoRA to improve model performance. These methods, though accurate, were  
Page 3319  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
costly and required powerful infrastructure.  
From these studies, it was clear that there is a need for a system that combines local deployment, secure  
authentication, and retrieval-based intelligence. The proposed AI-Driven Internal Data Intelligence Assistant  
addresses this by blending the privacy ideas of DB-GPT, the secure access concepts of Confidential  
Prompting, and adaptable enterprise fine-tuning — all under a RAG-based setup using ChromaDB, Ollama,  
and Supabase to ensure security, accuracy, and scalability.  
Existing solutions:  
Numerous enterprise-grade solutions that help document analysis and knowledge search have been developed  
in response to the growing demand for incorporating artificial intelligence into corporate workflows [4][5].  
These systems typically make an effort to strike a balance between the strict data governance requirements of  
contemporary businesses and the potent capabilities of large language models (LLMs).  
Microsoft Copilot for Microsoft 365:  
It is a highly integrated assistant Microsoft Copilot for Microsoft 365, which functions inside the Microsoft  
ecosystem. The company's "Commercial Data Protection" guarantee, which guarantees that organizational data  
and customer prompts are kept separate within the tenant's compliance boundary and aren't used to train  
foundational models, as the company claims. It offers a smooth but vendor-locked experience by using the  
Microsoft Graph to contextualize responses in user-specific emails, documents, and calendars. This makes the  
service very expensive hence small organizations cannot always afford this system.  
Glean:  
An enterprise search and discovery platform powered by AI, creates a unified, searchable knowledge graph  
that links together a company's diverse data sources ranging from Google Drive and Confluence, to GitHub  
and Salesforce. Glean adheres to existing access permissions, so users will only receive answers from the  
documents and data they already have access to. Regardless of these features, we must note that Glean is a  
cloud-based SaaS model and it is not suitable for domain specific industries.  
Bloomberg GPT:  
This offers a different solution. It is a domain specific model that has been trained from scratch on a very large  
body of financial data. This allows for maximizing performance and accuracy relative to the financial domain,  
and because the model is trained and deployed internally, they have access to the data, and thus full data  
sovereignty, minimizes privacy concerns regarding external APIs. The drawback lies in the trained model that  
is domain specific and cannot be reached for general corporate use.  
Even though these solutions can be innovative, they typically require substantial financial resources, force  
platform locking, or are only for specific industries. There is also no check-in for the data security that they  
might claim which cannot compete with the security against a model that only processes the data locally, in  
your own system and doesn’t export the data. Thus, there is no generalizable, affordable, and deployable  
solution in an effective manner for small to mid-sized businesses.  
Proposed Solution  
Our AI assistant is a standalone web-based solution that allows a company’s employees to interact with their  
corporate documentation in natural language. The system is designed using a "Retrieval First, Generation  
Second" construct inspired by RAG [1][3].  
Key Principles:  
1. Data Isolation: All company data remains inside the company's control infrastructure.  
2. Contextual Grounding: All AI generated content is grounded in retrieved content from the company's  
internal document store.  
Page 3320  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
3. Transparency: The system provides citations, referencing all documents used to generate every  
answer.  
Main Components:  
1. Document Ingestion Component: Accepts and pre-processes various document formats (PDF, DOCX,  
TXT).  
2. Vector Embedding and Storage Component: Embeds text into numerical representations  
(embeddings) and stores in a vector database.  
3. Query Processing and Retrieval Component: Understand user intent and retrieves the most  
semantically relevant text chunks. The assistant responds based on retrieved context and prior  
interactions instead of follow up questions.  
4. Response Generation Component: Generates a natural language content answer based on retrieved  
context only.  
Advantages Of Proposed Solution  
The proposed project “AI driven Internal Data Intelligence Assistant” brings advantages over existing  
solutions [4][5] and as a stand-alone system that operates independently, unrestricted by dependency on data  
sharing with public tools.  
1. By using a local LLM inference engine (Ollama), our system refuses to compromise the company data.  
This makes the system fit perfectly with organisations with sensitive data or security conscious values.  
2. The project uses open-source technologies and doesn’t use any paid API or services from the web,  
making the solution highly cost effective and affordable. This way we can serve small organisations by  
providing economic advantage.  
3. The proposed project supports the elimination of vendor lock-in by using open-source technologies like  
python, Ollama, ChromaDB. Hence the system is not exclusively tied to any specific vendor, helping the  
system gain flexibility and independent operations.  
4. The system architecture itself provides for the high relevance and accuracy in outputs from the assistant.  
This is because the system AI’s knowledge context is based around company data itself, hence the  
system delivers answers that provide high accuracy and relevance. Hence rated higher against public  
tools that are inflicted with noise, hallucination and unnecessary generic information.  
Working Details  
1. At the center of this system is something called a Retrieval-Augmented Generation (RAG) pipeline —it  
is a smart way to make AI give accurate answers using real information instead of random guesses.  
2. Here’s what actually happens behind the scenes: First, all the documents are cleaned up and broken into  
smaller, meaningful pieces (usually around 500–1000 characters long) so they’re easier for the AI to work  
with. Then, each of these pieces is turned into a special kind of digital “fingerprint” using a model like  
Sentence-BERT.  
3. These fingerprints are stored in a vector database (like ChromaDB), which is designed to quickly find  
text pieces that have similar meanings. Hence, when someone asks a question, that question is also  
converted into a fingerprint, and the system searches for the few most relevant text chunks — usually the  
top 3 to 5.  
Page 3321  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
4. Once those pieces are found, they’re placed into a prompt that basically says: “Use only this information  
to answer the question.” The whole thing is then sent to a Large Language Model (LLM), which reads  
the context and writes a clear, accurate answer — as if it’s pulling the answer straight from the documents.  
System Architecture  
Fig 1 - System Architecture Diagram  
We aim to build a system that is capable of finding relevant and accurate information effortlessly — a place  
where users can simply ask questions in natural language and get results pulled straight from their own  
documents.  
User Interaction: There are two types of users that contribute to the system. Regular employees, who are  
direct beneficiaries of the project. They can use queries and ask questions to the assistant; and admins, who  
handle uploading and managing the documents. Both use the same web dashboard — a single space that works  
like a chat and a control center rolled into one.  
Frontend Layer: The web interface is responsible for direct interaction of the user with the system. It is  
simple and intuitive — users log in, upload files, type in their questions, and instantly get answers. Behind the  
scenes, it quietly manages authentication, routes every request to the right service, and brings the AI’s response  
back to the user.  
Backend Services: Backend controls the system working, manages the logic, data, and server-operations. The  
backend is split into three main parts:  
Page 3322  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025  
1. Auth Service makes sure everything stays secure. It checks every user’s identity using JWT tokens and  
role-based access control before letting them in.  
2. Document Ingestion takes care of the uploaded PDFs. It extracts all the text, splits it into smaller  
chunks, generates metadata, and stores everything neatly in MongoDB — from the raw documents to  
chat history.  
3. The Query Processor is the system’s decision-maker. When a user asks a question, it searches for  
relevant information from the vector database, prepares the context, and coordinates with the AI model to  
generate a meaningful, accurate answer.  
Data Layer: Each database is responsible for specific tasks in our system.  
1. PostgreSQL manages the structured stuff — like user accounts and system logs.  
2. MongoDB handles the flexible data of the system— chat records, document logs, etc.  
3. ChromaDB stores the document embeddings — the numerical “understandings” of text that allow the  
system to match meaning, not just exact words.  
LLM Engine: The LLM Engine that powers our system is Ollama. Ollama supports the language model to  
understand query and generate results. Once the Query Processor finds the right pieces of information, Ollama  
blends them with the user’s question and generates the response that is completely based on the documents  
(company data) uploaded to the system by the admins. Hence, this ensures that the results are accurate and  
relevant, and not degraded by unnecessary generic information.  
In simple terms: You upload your PDFs, the system processes and stores them intelligently, and later, when  
you ask something, it digs into your own data to give you the right answer — fast, reliable, and backed by  
facts.  
CONCLUSIONS  
This research offers a thorough explanation of the system “AI - Driven Internal Data Intelligence  
Assistant”, the application of RAG and in-depth functionality of the project. The paper explains how our  
project is an appropriate solution to resolve the conflict between inclusion of artificial intelligence to promote  
productivity and risking the company's data security. By using the RAG pipeline, we demonstrate how AI and  
open-source technologies benefit the users without being under control of public tool providers.  
The proposed solution successfully completes the necessary checks to make a closed-loop environment for  
data processing. The integration of Ollama as a local LLM inference engine is one of the most crucial  
foundations of our work. It holds the centre of the system architecture and ensures that no company data leaves  
the organisational firewall. The assistant provides the employee with a natural language-based interface that  
gives accurate and relevant results and values transparency.  
This paper confirms that development of secure corporate tools is needed and can be built with design  
explained in this article. The project has been carefully developed in accordance with industry standards.  
Future work will include the expansion of features in terms of scalability, role-based access control, optimizing  
the document retrieval mechanism, using image/video-based information uploads, and addition of multi-hop  
queries. This project stands as evidence to the idea that AI capabilities and data security can coexist.  
REFERENCES  
1. Xue, L., Chen, M., and Li, Y., “DB-GPT: Empowering Database Interactions with Private LLMs,” IEEE  
International Conference on Data Engineering, 2024.  
2. Zhang, Y., Wu, H., and Zhao, T., “Confidential Prompting: Privacy-Preserving LLM Inference on Cloud,”  
IEEE Conference on Secure and Trustworthy Machine Learning, 2023.  
3. Kumar, R., and Singh, A., “E2E Data Extraction Framework from Unstructured Data: Integration of Deep  
Learning and Text Mining Techniques,” Journal of Intelligent Information Systems, 2025.  
4. Raj, S., Mehta, P., and Bansal, V., “Fine-Tuning Large Language Models for Enterprise Applications,”  
IEEE Conference on Artificial Intelligence and Data Science, 2024.  
Page 3323  
www.rsisinternational.org