🧠 Dynamic AI Customer Support Backend

An end-to-end AI-powered customer support backend built with FastAPI, combining offline ingestion, vector search (FAISS), intent detection, human feature extraction, LLM reasoning, and response strategy selection.

This system is designed using clean modular architecture, separating offline processing and online query execution for scalability and maintainability.

---

🚀 Key Features

• 🔹 Offline document ingestion & preprocessing
• 🔹 Chunking + embeddings using Sentence Transformers
• 🔹 FAISS-based vector similarity search
• 🔹 Intent & emotion detection
• 🔹 Human behavior feature extraction
• 🔹 Context-aware retrieval routing
• 🔹 LLM-powered response generation
• 🔹 Answer validation to reduce hallucinations
• 🔹 Strategy-based response selection
• 🔹 FastAPI REST interface

---

🏗️ System Architecture Overview

Offline Timeline (One-Time / Batch Process)

1. Load raw text data
2. Clean & preprocess documents
3. Chunk large documents
4. Generate embeddings
5. Enrich metadata
6. Store vectors in FAISS index

Online Timeline (Per User Query)

1. Preprocess user query
2. Extract human behavior features
3. Detect intent & emotion
4. Route retrieval strategy
5. Retrieve relevant chunks
6. Assemble contextual prompt
7. Generate response using LLM
8. Validate answer confidence
9. Return final response

---

📁 Project Structure

backend/
├── app/
│   ├── main.py                # Application entry point
│   ├── __init__.py
│
│   ├── ingestion/             # Offline ingestion pipeline
│   │   ├── data_load.py
│   │   ├── preprocessing.py
│   │   ├── embedding.py
│   │   ├── metadata_enricher.py
│   │   ├── ingestion_manager.py
│   │   ├── run_preprocessing.py
│   │   └── __init__.py
│
│   ├── intent_detection/      # Intent & emotion detection
│   │   ├── intent_classifier.py
│   │   ├── intent_features.py
│   │   └── __init__.py
│
│   ├── query_pipeline/        # Online query processing
│   │   ├── query_preprocess.py
│   │   ├── human_features.py
│   │   ├── query_embed.py
│   │   ├── context_assembler.py
│   │   ├── retrieval_router.py
│   │   └── __init__.py
│
│   ├── vector_store/           # Vector storage layer
│   │   ├── faiss_index.py
│   │   └── __init__.py
│
│   ├── reasoning/              # LLM reasoning
│   │   ├── llm_reasoner.py
│   │   ├── response_generator.py
│   │   └── __init__.py
│
│   ├── response_strategy/      # Response style selection
│   │   ├── response_router.py
│   │   ├── response_strategy.py
│   │   └── __init__.py
│
│   ├── validation/             # Answer validation
│   │   ├── answer_validator.py
│   │   └── __init__.py
│
│   └── data/
│       └── training_data.txt   # Knowledge base

---

🧩 Core Components Explained

🔹 Ingestion Pipeline (ingestion/)

Handles offline data preparation:

• Reads large text files
• Cleans & chunks content
• Generates embeddings
• Enriches metadata
• Prepares data for vector storage

🔹 Intent Detection (intent_detection/)

Detects:

• User intent (greeting, question, complaint, etc.)
• Emotional tone (angry, neutral, urgent)

🔹 Query Pipeline (query_pipeline/)

Online query execution:

• Cleans user input
• Extracts human behavioral features
• Embeds queries
• Retrieves relevant context

🔹 Vector Store (vector_store/)

• FAISS-based similarity search
• Efficient nearest-neighbor lookup

🔹 Reasoning Engine (reasoning/)

• Uses LLM to generate answers from retrieved context
• Applies system prompts dynamically

🔹 Response Strategy (response_strategy/)

• Chooses response tone (polite, empathetic, concise, etc.)
• Adjusts based on intent & emotion

🔹 Validation (validation/)

• Ensures answers are grounded in context
• Reduces hallucinations via confidence scoring

---

⚙️ Tech Stack

• Backend Framework: FastAPI
• Embeddings: Sentence Transformers (all-MiniLM-L6-v2)
• Vector Search: FAISS
• LLM: TinyLlama 1.1B Chat
• Data Processing: NumPy
• API Schema: Pydantic

---

▶️ Running the Application

1️⃣ Install Dependencies

pip install -r requirements.txt

2️⃣ Set Training Data Path

Update in app/main.py:

DATA_PATH = "/path/to/training_data.txt"

3️⃣ Start the Server

uvicorn app.main:app --reload

4️⃣ API Endpoints

• Health Check

GET /

• Query Chatbot

POST /query
Content-Type: application/json

{
  "user_query": "How do I reset my password?"
}

---

🧪 Example Response Flow

1. User sends a query
2. Intent + emotion detected
3. Context retrieved from FAISS
4. LLM generates response
5. Validator checks confidence
6. Final answer returned

---

📌 Future Improvements

• Streaming responses
• Multi-language support
• Persistent session memory
• Redis-based caching
• Async ingestion
• Hybrid search (BM25 + vectors)

---

👨‍💻 Author

Jenish Shekhada AI Engineer | GenAI | RAG Systems | FastAPI