🔍 Semantic Search API using FAISS & Sentence Transformers

This project is a simple Semantic Search API built using FastAPI, Sentence Transformers, and FAISS. The primary goal is to provide a practical understanding of how AI models find meaning-based similarities between sentences—a foundational concept behind advanced systems like Retrieval-Augmented Generation (RAG) and ChatGPT search.

---

🧠 What is Semantic Search?

Semantic Search is the process of finding text that has a similar meaning rather than relying on exact keyword matches.

For example:

The query “Doctor helps patients” will be considered semantically similar to the stored document “Physicians treat people.”

This mechanism is achieved using vector embeddings, which mathematically convert text into high-dimensional numerical representations of their meaning.

---

🧩 How Sentence Transformers & FAISS Work Together

Semantic Search relies on two core components:

Component	Purpose
Sentence Transformers	Converts text/sentences into embeddings (dense vector numbers) that capture semantic meaning.
FAISS (Facebook AI Similarity Search)	A highly efficient library used to store and rapidly search through these embeddings to find semantically similar entries.

Example Workflow:

1. Encoding: Encode the source text sentences using Sentence Transformers.
2. Indexing: Store the resulting embeddings inside a FAISS index.
3. Searching: When a new query is entered, it is first encoded, then searched against the FAISS index to get the most similar results instantly.

---

⚙️ Technologies Used

Library	Purpose
fastapi	For building and serving the high-performance REST API.
uvicorn	ASGI server for running the FastAPI application.
pydantic	For data validation of request and response models.
sentence-transformers	For generating meaning-based text embeddings.
faiss-cpu	For fast nearest-neighbor (vector similarity) search.
google-genai	For text generation and integration with the Gemini LLM.
python-dotenv	For securely loading environment variables like API keys.

💡 Note: You can switch faiss-cpu to faiss-gpu if your system supports CUDA for accelerated vector search.

---

🧰 Installation & Setup

1️⃣ Clone the repository

    git clone https://github.com/Aashish2k1S/Semantic_Search_API.git
    cd Semantic_Search_API

2️⃣ Create virtual environment & install dependencies

    pip install -r requirements.txt

3️⃣ Add your Gemini API key

Create a .env file in the root folder of the project:

    GEMINI_API_KEY=your_api_key_here

4️⃣ Run the API

    uvicorn main:app --reload

The API documentation will be live at → http://127.0.0.1:8000/docs

🧩 Example Request

Endpoint: POST /search

Request Body:

    {
        "question": "How does AI help doctors?"
    }

Example Response:

    {
        "question": "How does AI help doctors?",
        "context": [
            "AI is transforming healthcare and diagnostics.",
            "Deep learning assists in image-based medical analysis.",
            "FastAPI enables rapid API development for AI apps."
        ],
        "answer": "AI is transforming healthcare and diagnostics, offering significant assistance to doctors. It helps by revolutionizing the way medical information is processed and understood. Specifically, deep learning, a key AI technology, assists doctors with medical analysis. This support is particularly beneficial for image-based medical analysis, aiding in more accurate and efficient diagnoses."
    }

---

💡 Learning Goals

This mini-project served to solidify the following concepts:

• What embeddings are and how text can be represented numerically.
• The function of FAISS in performing rapid similarity search.
• How to expose an ML workflow through an API using FastAPI.
• Understanding the core foundation of RAG (Retrieval-Augmented Generation).

---

🧩Future Plans

• Add real dataset ingestion (e.g., text files, articles, PDF content) instead of mock data.
• Integrate the Gemini or OpenAI LLM to generate a context-based answer from the retrieved documents.
• Implement saving and loading of the FAISS index for persistent storage.
• Deploy the application on platforms like Render or Hugging Face Spaces.

---

👨‍💻 Author

Aashish Gupta
.NET & Python Developer | Exploring AI Systems

🔗 LinkedIn | GitHub

---

🌱 “I’m not chasing perfection — I’m learning by doing, one project at a time.”