Ending the Search Bottleneck: Custom RAG Chatbot Development for an EdTech Leader

NIX partnered with an EdTech company to develop a custom, agentic RAG chatbot that transforms years of scattered Slack logs into a secure, searchable corporate brain.

Business Domain

Education
Service

AI, AI Agent
Technologies

LangChain, LangGraph, FastAPI, Gradio, AWS App Runner

Business Overview

Our client is a U.S.-based EdTech company offering a medical exam preparation platform with high-fidelity mock exams and learning tools. As critical knowledge became scattered across years of unstructured Slack conversations, employees had to spend hours searching disconnected channels and lengthy threads to find past decisions.

To eliminate these inefficiencies and preserve institutional knowledge, the client set out to build an AI-powered assistant capable of instantly retrieving accurate historical information through natural conversation. Following a successful previous collaboration, they partnered with NIX to design and develop this custom Slack RAG chatbot (retrieval-augmented generation) from the ground up.

Project Scope

End-to-end system architecture: Designing the structural blueprint for data flow, routing mechanisms, and storage components.
Data pipeline development: Architecting the ingestion frameworks to programmatically clean, restructure, and vectorize internal communications.
Core agentic engine implementation: Building the chat agent’s logic layer, routing mechanisms, and backend API.
User interface integration: Creating an accessible, interactive frontend workspace for internal testing and evaluation.
Cloud infrastructure deployment: Provisioning scalable, containerized environments within the cloud ecosystem to host the complete application.

Challenge

The client needed a secure agentic RAG chatbot delivered within just two months, leveraging cost-effective AWS-hosted LLMs to manage expenses without compromising output quality. This created several key challenges:

Limited model capabilities required advanced prompting techniques to achieve reliable results.

Complex, unstructured Slack conversations had to be transformed into searchable, well-organized knowledge.

Strict accuracy requirements demanded precise PII redaction, noise removal, and thread restructuring without introducing AI hallucinations.

Solution

To deliver a precise knowledge retrieval agent within a strict timeline, NIX bypassed standard, single-pass chatbot designs. Instead, we engineered a predictable, state-driven, lightweight agentic architecture that utilizes the LLM as a controlled transformation component within a highly structured workflow.

Iterative Ingestion and Intelligent Data Transformation Pipeline

The foundation of the platform is a sophisticated, chronological preprocessing pipeline built from the ground up to clean and normalize raw, unstructured communication data.

Multi-step noise reduction and sanitization: Slack logs are extracted programmatically and passed through a custom cleaning pipeline. The system combines deterministic scripts with targeted LLM calls to execute noise reduction and perform both manual and automated PII redaction to protect internal credentials.
Normalized question-answer-summary format: To compensate for the lower reasoning capabilities of smaller LLMs, the pipeline handles data transformation iteratively. Chaotic chat threads are first synthesized into a highly structured question-answer-summary format.
Automated self-check validation loop: The output then undergoes a secondary validation pass that acts as a built-in quality gate. This loop audits the generated summaries for clarity, factual grounding, and standalone reusability, ensuring a high signal-to-noise ratio and completely minimizing hallucinations before storage.

Budget-friendly Architecture and Data Isolation

We designed the storage layer to keep data private, secure, and highly cost-effective.

Cost-saving storage design: Instead of setting up expensive, traditional database infrastructure during this testing phase, NIX built a custom S3-backed vector bucket using the client’s existing AWS account. This gave the client a budget-friendly, secure, and scalable repository that keeps costs remarkably low during the tool evaluation.
Complete data isolation: To ensure total privacy, the entire search system runs completely inside the client’s own cloud ecosystem. Valuable company insights and internal discussions remain 100% private, protected from outside eyes, and safely tucked away in an isolated digital storage.

Intelligent Query Routing and Accurate Search

The system is built to provide precise historical answers rather than generic AI guesses. Every time a user types a query, the platform instantly determines the absolute best path to retrieve the exact data needed.

Intent classification layer: Driven by LangGraph and LangChain, the chat agent utilizes an intent routing mechanism. When a user queries the system, the architecture instantly analyzes the input to decide the optimal reasoning path: determining whether the request requires a deep RAG-based archive search, a general knowledge response, or a simple conversational interaction.
Semantic retrieval: For historical queries, the user’s question is vectorized in real-time. The system scans the S3 vector bucket to extract the top 5 most semantically similar text blocks. This precise context is then fed directly into the core LLM, allowing it to generate a highly precise, fact-grounded answer based strictly on verified historical company decisions.

User-friendly App Design and Fast Cloud Setup

The system was built for fast testing, easy day-to-day use, and effortless future updates.

Simple, intuitive chat workspace: The tool features a clean, user-friendly chat interface backed by high-performance APIs. This gives the client’s internal team a smooth, hassle-free environment to test features, ask questions, and evaluate results.
Secure and hassle-free cloud hosting: The entire application is packaged safely and hosted directly on AWS. This gives the client a private, stable, and highly secure testing environment that is easy to maintain and ready for simple model upgrades as AI technology evolves.

Outcome

This project successfully achieved its main objective, serving as a high-performing PoC that demonstrated the viability and power of internal AI tools for the client’s business.
Following the success of this PoC, NIX continues to serve as a trusted AI partner, currently collaborating on higher-priority, client-facing commercial AI initiatives.

Eliminated knowledge rot: Transformed years of unorganized chat logs into an instant, searchable corporate brain, saving employees hours of manual scrolling and preserving vital institutional memory.
Rapid time-to-value: Designed, built, and deployed a fully functional solution completely from scratch in under two months, allowing the client to evaluate AI capabilities with zero operational delay.
High ROI on a lean budget: Achieved high answer accuracy, minimizing infrastructure expenses without sacrificing search precision.
100% data privacy: Ensured complete data isolation within the client’s own cloud ecosystem, protecting sensitive internal communications from public LLM exposure.

Team:

3 experts ( AI engineer, Tech Lead, Project Manager )

Tech stack:

FastAPI, LangChain, LangGraph, Gradio, AWS App Runner

Relevant Case Studies

View all case studies

Scaling Medical Exam Generation With AI

Education

65% Less Manual Work: Healthcare Document Automation With AI

Healthcare

AI-powered Search Solution for a Healthcare Company

Healthcare

AI Chatbot for Personalized Healthcare

Healthcare

vSentry—AI Web App for Vehicle Monitoring

Cybersecurity

Electronics

Success Story vSentry—AI Web App for Vehicle Monitoring image

Voice AI in Healthcare: Faster Specialty Medication Access

Healthcare

Contact Us

Processing…

I am interested in:
- Software solutions and services
- PR & Media
- Partnership
- Career
- Other
First Name**
Last Name**
Phone Field Validation
Phone number**
Email**
Leave us a message**
Attach file
Accepted file types: jpg, jpeg, png, webp, pdf, docx, txt, Max. file size: 50 MB.
*
I have read and I agree to the Privacy Policy
*
I have read and agree NIX United’s online cookie notice
This field is hidden when viewing the form
Phone Field Full Number
This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.