Semantic Search Engine (SSE)

A FastAPI-powered semantic search application with React frontend, featuring document uploads, vector embeddings, and LLM-powered question answering.

✨ Features

• Document Processing: Upload PDFs, Word docs, and text files
• Semantic Search: Find relevant content using vector embeddings
• RAG Integration: Generate AI-powered answers from your documents
• Session Management: Isolate document sets per session
• Modern Stack: FastAPI + React + FAISS + HuggingFace models

Feature	Description
Multi-format Support	PDF, Word, Text files
Hybrid Search	Semantic + keyword search
RAG Integration	Context-aware answers
Performance	Optimized chunking & indexing

🛠️ Installation

Docker runner

docker compose up --build
# Accsess the app at http:localhost:8000

Development Setup

   # Linux / Mac
   python -m venv venv
   source venv/bin/activate 

   # Windows
   venv\Scripts\activate

   # run
   pip install -r requirements.txt
   uvicorn main:app --reload

Frontend

   cd frontend
   pnpm install
   pnpm run dev

API Endpoints

Endpoint	Method	Description
/api/session	GET	Create a new session
/api/upload	POST	Upload files and resource
/api/search	GET	Perform semantic search

How the app operates?

1. You Open the App

   The app creates a unique session ID for you

   Behind the scenes:
   → Generates a random ID
   → Creates an empty folder named after that ID to store your files.

2. You Upload Files

   You drag/drop PDFs, Word docs, or text files into the app.

   Behind the scenes:
   → Saves files to your session folder.
   → Breaks each file into small text chunks (e.g., 1-2 sentences each).
   → Stores these chunks in a list with metadata (file name, page number, etc.).

3. The App "Understands" Your Files

   The app converts every text chunk into number sequences (vectors) using AI.

   Behind the scenes:
   → Uses a pre-trained model (all-MiniLM-L6-v2) to generate vectors.
   → Builds a searchable index using FAISS (Facebook’s search tool).

4. You Search for Something

   Behind the scenes:
   → Converts your query into a barcode (vector) using the same AI model.
   → Compares it against all document barcodes to find the closest matches.
   → Reranks results using a second AI (cross-encoder) to prioritize relevance.

5. You Get Results

   The app shows you:
   -> Direct excerpts from your documents (sorted by relevance).
   -> AI-generated summary (if RAG is enabled), combining the top matches into a natural answer.

Simple Analogy

I. Session ID = Your private locker.

II. File Upload = Putting books into the locker.

III. Chunking = Tearing out pages and highlighting paragraphs.

IV. Vectors = Giving each paragraph a unique ID number.

V. Search = Finding paragraphs with matching ID numbers to your question.

VI. AI Answer = A friend (the LLM) reads those paragraphs and explains the answer to you.