Jagruti System Design

Architecture

The system has three core models:

• Chatbot – Interacts with the user, interprets queries, and provides answers.
• Embedding Model – Converts documents into vector embeddings for semantic search.
• Summariser – Generates structured metadata and concise descriptions for new documents.

We use ChromaDB to store all document embeddings along with their metadata.
ChromaDB enables fast similarity search and supports filtering by metadata (e.g., category, date, author).

---

Workflow

1. Document Creation & Storage

1. A new document is uploaded to the intranet.
2. Summariser generates structured metadata:
   • Document name
   • Date of creation
   • Date of last modification
   • Category (e.g., Faculty Data, Student Data, Hostels, Academics, Messes)
   • A brief description of the content
3. Embedding Model processes the document:
   • Tokenises and chunks the text into manageable sizes.
   • Generates vector embeddings for each chunk.
4. The embeddings and their metadata are stored in ChromaDB.

---

2. User Query & Response

1. The user submits a question via the website.
2. Chatbot:
   • Converts the query into an embedding.
   • Uses ChromaDB to retrieve the most relevant chunks, filtered if needed by metadata.
   • Reads the retrieved content and composes a natural-language answer.
3. The interface displays:
   • Chatbot’s answer.
   • Highlighted Matches – exact sentence/phrase from the matched chunk.
   • References to relevant documents.

---

3. Website & API Layer

Frontend (Next.js)

• Responsive UI for desktop and mobile.
• User authentication and session management.
• Chat interface for queries and responses.
• Metadata-based filtering (category, date range, author, “recently added”).
• Quick Filters & Facets for faster narrowing of results.
• Keyboard Shortcuts – direct typing focuses chatbox.
• Document preview thumbnails (PDF, text, image icons).
• Clickable links to intranet documents.

Backend API (FastAPI)

• Acts as a bridge between frontend, ChromaDB, and AI models.
• Endpoints for:
  • Document upload
  • Metadata retrieval
  • Search queries
  • Embedding creation
  • Summarisation
  • Chatbot queries
• Returns JSON for integration with Next.js.
• Handles filtering and sorting before passing to chatbot.

Integration

• Next.js frontend ↔ FastAPI backend via REST.
• FastAPI interacts with ChromaDB and AI models.
• Authentication via Institute CAS for secure access.

---

Admin Panel

Accessible only via CAS authentication.

Features:

• Document upload form with metadata fields.
• Upload Progress & Status – from upload → summarising → embedding → ready.
• List of uploaded documents with:
  • Status
  • Last update time
  • Quick re-upload option

---

Database Schema

ChromaDB is metadata-driven. We maintain two main collections:

1. documents (metadata only, 1 row per document)

Field Name	Type	Description
doc_id	UUID	Unique ID for the document
name	String	Document name
category	Enum	One of: faculty, student, hostel, academics, mess
description	String	Summary from Summariser
created_at	DateTime	Document creation date
updated_at	DateTime	Last modification date
file_url	String	Link to intranet file
author	String	Optional uploader’s name

2. chunks (embeddings + chunk metadata, 1 row per chunk)

Field Name	Type	Description
chunk_id	UUID	Unique ID for the chunk
doc_id	UUID	FK to documents.doc_id
chunk_text	Text	Actual chunk text
embedding	Vector[float]	Vector representation from Embedding Model
position	Int	Chunk order in document
category	Enum	Same as document category

---

ChromaDB Notes

• embedding is stored in the vector index for similarity search.
• Metadata fields like category and doc_id are indexed for filtering.

---

Models

We are using:

• Qwen 3-0.6B via Hugging Face Text Generation Inference for chat and summarization
• Qwen/Qwen3-Embedding-0.6B via sentence-transformers for embeddings
• pdfplumber for PDF text extraction (replacing PyPDF2)

Technical Implementation

PDF Processing

• Uses pdfplumber for robust PDF text extraction
• Handles complex PDF layouts better than PyPDF2
• Preserves text formatting and structure

Embedding Generation

• sentence-transformers with Qwen-based embedding model
• Generates high-quality vector representations for semantic search
• Embeddings are normalized for better similarity calculations

Chat & Summarization

• Hugging Face TGI with Qwen 3-0.6B model for natural language tasks
• Handles document summarization during upload
• Provides contextual responses using retrieved document chunks

Configuration

Copy .env.example to .env and configure:

• HF TGI server URL and model settings
• Embedding model preferences
• ChromaDB storage location
• File upload limits and allowed types

Assumptions

• All data must be machine-readable for LLMs.
• Images require alt-text or descriptive metadata.
• Handwritten text must be OCR-processed before upload.
• Only supported formats (PDF, plain text, etc.) are allowed.