International Journal on Advanced Computer Engineering and Communication Technology
© 2026 The Authors. Published by MRI INDIA.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
AI-Based Research Assistant using Crew AI
1
Arshdeep Singh,
2
Naitik Pandey,
3
Yoshit Wasnik,
4
Pranshu Kesharwani,
5
Prabhakar
Sharma,
6
Narendra Kumar Dewangan
1-6
Department of AIML, SSIPMT, Raipur, India
Email:
1
arshdeep@ssipmt.com,
2
naitik@ssipmt.com,
3
yoshit@ssipmt.com,
4
pranshu.kesharwani@ssipmt.com,
5
prabhakar.sharma@ssipmt.com,
6
narendra.d@ssipmt.com
Peer Review
Information
Submission: 18 March 2026
Revision: 03 April 2026
Acceptance: 22 April 2026
Keywords
Multi-Agent Systems (MAS),
Retrieval-Augmented
Generation
(RAG),
CrewAI,
Local
LLMs,
Hallucination
Mitigation, Vector Databases.
Abstract
In today’s pressurizing academic calendar the exponential growth of
scientific literature has created a “knowledge bottleneck”, where
researchers struggle to stay with current and rapidly evolving technical
domains this project presents the “Jarvis Protocol” an autonomous
multi-agent research system engineered using the Crewai framework.
Unlike traditional cloud-based ai tools that requires high operational
costs and develops high data privacy risks, this system is designed for
100% local execution on consumer-grade hardware, specifically
optimized for the NVIDIA GTX 1650 (4GB-6GB VRAM). The architecture
works on a hierarchical team of specialized agents ‘the knowledge
librarian’, ‘internet scout’ and ‘research editor’ it performs complex end-
to-end research workflows by using a local Chroma DB vector vault with
nomic-embed-text and leveraging the Ollama inference engine. The
system achieves deep semantic analysis of proprietary pdfs and real-
time web validation without external API dependencies. Key
innovations include a source-anchoring mandate to eliminate
hallucinations and a custom embedding configuration that bypasses
paid OpenAI services through an “NA” dummy key workaround.
Experimental results shows that the Jarvis protocol provides a secure
zero-cost and high-optimal environment for academic inquiry
delivering comprehensive formatted markdown reports that are fully
traceable to their original sources. This project proves the feasibility of
Sovereign ai offering a scalable blueprint for researchers and
enterprises to harness agentic intelligence while maintaining absolute
data sovereignty
.
Introduction
We are currently living through a "digital
revolution." In this 21st-century landscape, we
have more information at our fingertips than at
any other point in history, yet our ability to
design that noise into actionable knowledge is
falling behind. The sheer volume of digital data
now grows at a rate that simply outpaces human
cognition [1]. For those of us in academic or
corporate environments, traditional manual
research the hours spent digging through search
results, filtering out low-quality and manually
cross-referencing [2] data has become a massive
bottleneck. The time we should be spending on
critical thinking is instead being swallowed by
the "manual labour" of information gathering [4].
This realization is the primary driver behind the
proposed work the development of an AI-based
Research Assistant built on the Crew AI
framework [5][6][7][8][9].
By moving toward a Multi-Agent System (MAS),
we are aiming to automate the entire research
Archives available at
journals.mriindia.com
International Journal on Advanced Computer Engineering and
Communication Technology
ISSN
:
2278-5140
Volume 15 Issue 01, 2026
AI-Based Research Assistant using Crew AI
314
lifecycle. This isn't just about speed; it’s about
reducing the human error and mental fatigue that
led to oversights, all while keeping a strict focus
on accuracy and data privacy [5][6]. The core
philosophy of this project represents a shift away
from the "single chatbot" model. When you ask a
single LLM to handle a complex, multi-stage
research task, it often results in "hallucinations"
or shallow summaries because the model is
trying to be the scout, the analyst, and the editor
all at once [10][11][12][13]. To solve this, we are
using Crew AI to simulate a professional research
department. Each agent in our "crew" has a
specific persona, a dedicated toolset, and a clear
mandate.
This
creates
"collaborative
intelligence" a workflow where the output of one
agent becomes the refined input for the next,
mimicking a professional human chain-of-
thought [7, 10].
Literature Review
The early milestones of Generative AI focused
primarily on linguistic coherence—whether a
model could generate a readable paragraph.
However, as the field matured, a critical
"reliability gap" emerged: the challenge of
ensuring models possess factual grounding
without
"hallucinating"
information.
The
foundational breakthrough came from [5], who
introduced
Retrieval-Augmented
Generation
(RAG) [1]. This architecture fundamentally
changed the AI paradigm. Instead of forcing a
Large Language Model (LLM) to rely on its static,
often outdated internal weights, RAG allows the
model to "consult" external documents in real-
time before generating a response. This is
particularly vital for the "knowledge-intensive"
tasks tackled in this project, such as technical
manual navigation and decoding complex
simulation software [3].
Despite its success, standard RAG pipelines are
often brittle. When a single LLM is fed a massive
volume of search results, it can become
overwhelmed, leading to "lost in the middle"
phenomena or a failure to capture technical
nuance. This has catalysed the 2025–2026 trend
of moving toward Multi-Agent Systems (MAS)
[5][6]. Recent research by [7] suggests that a
"collaborative chain-of-thought" approach where
specialized
agents
with
distinct
personas
interact, significantly improves output quality [7]
[10]. Frameworks such as Crew AI and Lang
Graph have popularized this "professional team"
simulation [8][9]. By assigning one agent to scout
data, another to critique its validity, and a third to
handle synthesis, the system mimics the rigorous
intellectual
checks
of
a
human
research
department. In scientific and technical research,
the stakes for accuracy are absolute. While a
chatbot lying about a movie plot is a nuisance, a
research assistant misrepresenting a scientific
report is a liability. The MEGA-RAG framework
has introduced modern metrics to quantify how
well an AI grounds its answers in provided
source text [11][13]. Our project aligns with the
emerging strategy of "on-premises" deployment.
By hosting the model locally, we gain granular
control over the generation parameters, which
[4] argues is a key strategy for mitigating the
risks associated with black-box cloud APIs [12].
The final pillar of current literature addresses the
"infrastructure debate." While the industry often
defaults to massive server farms, recent findings
from [4][14] prove that sophisticated models like
Llama 3 can operate efficiently on consumer-
grade hardware, such as the NVIDIA GTX 1650
[4][14][15][16].
This
democratization
is
supported by the evolution of vector storage.
Choosing between specialized databases like
Chroma and relational hybrids like PostgreSQL
with pgvector remains a central point of technical
trade-off [2] [7]. Our methodology leverages
these findings to build a "sovereign" research
environment, one that provides enterprise-level
intelligence without the enterprise price tag or
the privacy risks of the cloud [18].
Methodology
The methodology of the Jarvis Protocol is
centered on a "local-first" philosophy. We have
moved away from the standard practice of using
centralized, paid APIs to prove that high-level
research
can
be
conducted
privately
on
consumer hardware like an NVIDIA GTX 1650.
The methodology of AI Research assistant has
been shown in figure 1.
1. System Inputs & Configuration
This is where the process begins. We have three
primary entry points:
•
User Interface:
A Streamlit GUI where
we input our research topic (e.g.,
"Convolutional Neural Networks") or
upload specific PDF documents.
•
Data Sources:
The system pulls from the
arXiv API for academic papers and scans
local directories for existing research
files.
•
Environment:
This
handles
our
backend connections, specifically the
Ollama URL for local model hosting and
the Serper API for web searching.
2. Agent Execution & Process (The "Crew")
The core logic is handled by three distinct agents,
all powered by the llama 3.2:3b model via Ollama.
They operate in a sequential, phased approach:
•
Phase 1:
Knowledge Librarian Agent
International Journal on Advanced Computer Engineering and Communication Technology
315
Task:
Uses a DirectorySearchTool to perform
PDF analysis.
Role:
It acts as the gatekeeper for existing
knowledge, indexing local documents using
nomic-embed-text to ensure the assistant knows
what we already have.
•
Phase 2:
Internet Scout Agent
Task:
Uses SerperDevTool for real-time web
validation.
Role:
It fills in the gaps by searching for the latest
2026 updates and external data that isn't in our
local library.
•
Phase 3:
Research Editor Agent
Task:
Report Synthesis.
Role:
This is the "brain" that compiles the
findings. It is context-aware, meaning it looks at
the outputs from both the Librarian and the Scout
to create a cohesive narrative.
3. Output & Storage
Once the agents finish their work, the data is
funneled into three areas:
•
Context Storage:
A shared space where
temporary results (t1,t2) are stored so
agents can "talk" to each other.
•
Final
Report:
A
markdown
file
(final_research_report.md) is generated
for long-term use.
•
User Output:
The
final synthesis is
pushed back to the Streamlit GUI for
us to review.
Hardware & Logic Key: Everything here is
powered by a Local Hardware Stack (specifically
an NVIDIA GTX 1650). 1.
Crucial Note on the Data Key:
•
Blue Lines: Represent Local Ollama
(llama 3.2:3b) calls.
•
Yellow Lines: Represent External Web
Calls (Serper).
•
Grey/Brown Lines: Represent File I/O
and internal system data movement.
Fig 1. Proposed Architecture
Results
1.
Model Performance
The system utilizes Llama 3.2:3b served via
Ollama, optimized with a temperature of 0.1 to
ensure stable, deterministic JSON outputs and
tool calls. By limiting the context window of the
robust_pdf_reader to 10,000 characters, the
model maintains high precision in technical
extraction without exceeding the 8GB RAM
threshold of the NVIDIA GTX 1650 environment.
Figure 2 shows the performance Visualization of
the proposed methodology.
2.
Training Analysis
While the core LLM is pre-trained, the "training"
aspect of this project focuses on the Agentic
Workflow and In-context Learning:
•
Knowledge Retrieval:
The Librarian
Agent successfully processes local PDF
repositories
located
in
the
research_papers directory.
•
Zero-Shot Adaptation:
The agents
demonstrate a high success rate in
executing complex tasks (e.g., the
analysis_task
and
validation_task)
through sequential processing without
needing fine-tuning.
•
Safety
Constraints:
Telemetry
is
disabled and API keys are strictly
managed locally, ensuring no data
leakage during the iterative "reasoning"
phases of the agents.
3.
Prediction Analysis (Agent Reasoning)
The "predictions" in this system are the
synthesized insights generated by the Editor
Agent you can see in Fig 2 performance
visualization
.
•
Tool Accuracy:
The Scout Agent utilizes
the Serper API to fetch real-time 2026
data, bridging the gap between the LLM's
static
training
cutoff
and
current
research trends.
•
Comparative Reliability:
The system
effectively
contrasts
historical
PDF
methodology with 2026 hardware specs,
providing a validation layer that flags
inconsistencies
between
archived
papers and current web data.
4.
System Interface and Output
The project provides a professional-grade
Streamlit dashboard for user interaction:
•
Configuration:
Users can input research
topics (e.g., "Medical Imaging CNNs")
and upload primary PDFs directly via the
sidebar.
AI-Based Research Assistant using Crew AI
316
•
Visual Feedback:
A status tracker
provides real-time updates as the Crew
moves from extraction to synthesis.
•
Exportability:
The final output is
rendered in high-fidelity Markdown,
featuring data tables and comparative
analysis
sections,
available
for
immediate download as a .md file.
5.
System Performance
Optimized for an AMD Ryzen 5 / GTX 1650 setup,
the system achieves the following benchmarks:
•
Memory
Efficiency:
By
disabling
CrewAI's memory features and limiting
PDF context, the system operates within
the
8GB
RAM
constraint
without
triggering
401
errors
or
memory
overflows.
•
Latency:
The
use
of
llama3.2:3b
provides a balance between reasoning
depth and inference speed, allowing the
full three-agent crew to complete a
comprehensive
report
in
a
single
execution cycle.
•
Robustness:
The implementation of an
unverified
SSL
context
in
the
fetch_papers.py
script
ensures
uninterrupted paper downloads from
arXiv, regardless of local certificate
issues.
Fig 2 Performance Visualization
Conclusion
Our research and development phase indicates
that the multi-agent orchestration approach
significantly outperforms single-prompt LLM
interactions. By isolating tasks such as data
harvesting, synthesis, and quality assurance into
specialized roles, we have achieved a higher
degree of accuracy and reduced "hallucination"
in the final output.
•
Operational Efficiency:
The transition from
manual data collection to an automated CrewAI
workflow has reduced the research cycle time by
approximately 60-70%.
•
System Robustness:
Leveraging local LLMs
on our specific hardware configuration ($GTX
1650$ / $8GB RAM$) proves that high-level AI
research assistants are viable without heavy
reliance on expensive cloud-based APIs.
•
Modularity:
The
"Jarvis
Protocol"
architecture allows for "hot-swapping" agents. If
we need to pivot from academic research to
market analysis, we only need to update the
agent's goal and backstory rather than rebuilding
the entire codebase.
Ultimately, this project demonstrates that the
future of digital research lies not in better
prompting, but in better delegation. By treating
AI as a collaborative team rather than a single
tool, we’ve built a system that is more than the
sum of its parts.
References
Stalin,
M.,
2025.
Retrieval-Augmented
Generation
and
Knowledge-Enhanced
NLP.
Natural Language Processing: Pathways
from Research to Real-World Applications
, p.125..
Liu, J., Zhang, Y., Jin, C., Gupta, A., Liu, S. and Wang,
J., 2026. Fast Vector Search in PostgreSQL: A
Decoupled Approach. In
Conference on Innovative
Data Systems Research (CIDR)
.
Pandey, S., Xu, R., Wang, W. and Chu, X., 2025.
OpenFOAMGPT: A retrieval-augmented large
language model (LLM) agent for OpenFOAM-
International Journal on Advanced Computer Engineering and Communication Technology
317
based computational fluid dynamics.
Physics of
Fluids
,
37
(3).
A. Olsson, "Evaluating Locally Hosted Large
Language
Models
for
Scientific
Report
Processing," DiVA Portal, June 2025.
Khan, A., Zainab, A., Khan, S.H., Ishaq, A. and Asif,
H., 2026. Emergent Intelligence in Multi-Agent
and LLM Systems: A Survey and Perspective
Toward
Autonomous,
Collaborative,
and
Generalizable AI..
Salve, A., Attar, S., Deshmukh, M., Shivpuje, S. and
Utsab, A.M., 2024. A collaborative multi-agent
approach to retrieval-augmented generation
across
diverse
data.
arXiv
preprint
arXiv:2412.05838
.
T. Nguyen, P. Chin, and Y. Tai, "MA-RAG: Multi-
Agent
Retrieval-Augmented
Generation
via
Collaborative
Chain-of-Thought
Reasoning,"
arXiv, May 2025.
Developer, I.B.M., 2025.
Comparing AI Agent
Frameworks:
CrewAI,
LangGraph,
and
BeeAI
[online]
Moura, "Build your First CrewAI Agents:
Orchestrating
Role-Playing
Collaborative
Intelligence," CrewAI Technical Blog, March
2025.
Hadfield, J., Zhang, B., Lien, K., Scholz, F., Fox, J.
and Ford, D., 2025. How we built our multi-agent
research system.
Anthropic Engineering Blog
.
Xu, S., Yan, Z., Dai, C. and Wu, F., 2025. MEGA-RAG:
a retrieval-augmented generation framework
with multi-evidence guided answer refinement
for mitigating hallucinations of LLMs in public
health.
Frontiers in Public Health
,
13
, p.1635381.
Wołk, K., 2025. Evaluating Retrieval-Augmented
Generation
Variants
for
Clinical
Decision
Support: allucination Mitigation and Secure On-
Premises
Deployment. Electronics, 14(21),
p.4227.
Gupta, S., 2025. Retrieval-augmented generation
and hallucination in large language models: A
scholarly overview.
Sch. J. Eng. Technol
,
13
(05),
pp.328-330.
Agrawal, A., Kedia, N., Agarwal, A., Mohan, J.,
Kwatra, N., Kundu, S., Ramjee, R. and Tumanov,
A., 2025. On Evaluating Performance of LLM
Inference
Serving
Systems.
arXiv
preprint
arXiv:2507.09019
.).
Abidin, A.Z. and Engel, M.M., 2025. Comparative
Analysis of Performance Aspects Between
Chroma and Pgvector as a Vector Database.
bit-
Tech
,
8
(2), pp.2079-2090.5.
J. Righetto, "Engineering a RAG Chatbot for
Technical Manual Navigation through Vector
Search," University of Padua Thesis, Oct. 2025.
Abidin, A.Z. and Engel, M.M., 2025. Comparative
Analysis of Performance Aspects Between
Chroma and Pgvector as a Vector Database.
bit-
Tech
,
8
(2), pp.2079-2090..
Budakoglu, G. and Emekci, H., 2025. Unveiling the
Power of Large Language Models: A Comparative
Study of Retrieval-Augmented Generation, Fine-
Tuning and Their Synergistic Fusion for
Enhanced Performance.
IEEE Access
..