Intenta - Intelligent Intent Classification & Retrieval System

Overview

Intenta is an end-to-end Machine Learning system that classifies user intents and retrieves relevant domain-specific knowledge using a hybrid search architecture.

The project combines traditional NLP, semantic embeddings, vector search, and REST APIs to provide intelligent support-query understanding.

---

Live Demo

API Base URL: https://intenta-api-3r59.onrender.com/

Swagger Documentation: https://intenta-api-3r59.onrender.com/docs

Health Check: https://intenta-api-3r59.onrender.com/health

---

Example Workflow

User Query
    ↓
Language Detection
    ↓
Translation (if required)
    ↓
Text Preprocessing
    ↓
Intent Classification
(TF-IDF + Logistic Regression)
    ↓
Intent Routing
    ↓
Hybrid Retrieval
(FAISS + BM25)
    ↓
Relevant Response Documents

---

Features

Intent Classification

• TF-IDF Vectorization
• Logistic Regression Classifier
• Confidence-based Prediction
• Alternative Intent Suggestions

Hybrid Retrieval

• Semantic Search using FAISS
• Keyword Search using BM25
• Hybrid Ranking Strategy
• Domain-Specific Knowledge Bases

Multilingual Support

Supported Languages:

• English
• Hindi
• Gujarati

Queries are automatically translated to English before classification.

Analytics & Feedback

• User Feedback Collection
• Positive/Negative Feedback Tracking
• Intent Performance Analytics
• Failure Trend Monitoring

API Services

• Prediction API
• Semantic Search API
• Feedback API
• Analytics API
• Health Check API

---

Architecture

                    ┌─────────────────┐
                    │ User Query      │
                    └────────┬────────┘
                             │
                             ▼
                 ┌─────────────────────┐
                 │ Language Detection  │
                 └────────┬────────────┘
                          │
                          ▼
                 ┌─────────────────────┐
                 │ Translation Layer   │
                 └────────┬────────────┘
                          │
                          ▼
                 ┌─────────────────────┐
                 │ Text Cleaning       │
                 └────────┬────────────┘
                          │
                          ▼
                 ┌─────────────────────┐
                 │ Intent Classifier   │
                 │ TF-IDF + LogReg     │
                 └────────┬────────────┘
                          │
                          ▼
                 ┌─────────────────────┐
                 │ Intent Router       │
                 └────────┬────────────┘
                          │
                          ▼
          ┌──────────────────────────────────┐
          │ Hybrid Retrieval System          │
          │                                  │
          │ Semantic Search (FAISS)          │
          │ Keyword Search (BM25)            │
          └──────────────┬───────────────────┘
                         │
                         ▼
                 ┌─────────────────────┐
                 │ Retrieved Documents │
                 └─────────────────────┘

---

Tech Stack

Category	Technology
Backend	FastAPI
ML Framework	Scikit-Learn
Embeddings	Sentence Transformers
Vector Search	FAISS
Keyword Search	BM25
Language Detection	LangDetect
Translation	Deep Translator
Model Storage	Joblib
Data Storage	JSON
Evaluation	Scikit-Learn Metrics
Visualization	Matplotlib, Seaborn

---

Machine Learning Pipeline

Classification Model

TF-IDF Vectorizer
        +
Logistic Regression

Retrieval Model

SentenceTransformer
(all-MiniLM-L6-v2)
        +
FAISS

Hybrid Scoring

Final Score =
(0.7 × Semantic Score)
+
(0.3 × Keyword Score)

---

Dataset

Intent Dataset

• 11 Intent Classes
• Developer Support Queries
• Training Examples per Intent

Example Intents:

api_error
database_connection
deployment_issue
docker_issue
git_issue
login_issue
model_saving
payment_failed
project_run
project_transfer
training_error

Domain Knowledge Bases

technical
billing
account

---

API Endpoints

Method	Endpoint	Description
POST	/predict	Predict intent and retrieve documents
POST	/semantic/search	Semantic search endpoint
POST	/feedback	Submit user feedback
GET	/analytics	View feedback analytics
GET	/health	Health check

---

Predict Intent Example

Request

{
  "text": "my model training failed"
}

Response

{
  "prediction": "training_error",
  "confidence": 0.87,
  "source": "technical"
}

---

Results

Metric	Value
Accuracy	72.09%
Intent Classes	11
Languages Supported	3
Retrieval Method	Hybrid
API Framework	FastAPI

---

Running Locally

Clone Repository

git clone <repository-url>
cd Intenta

Create Virtual Environment

python -m venv venv

Activate Environment

Windows

venv\Scripts\activate

Linux / macOS

source venv/bin/activate

Install Dependencies

pip install -r requirements.txt

Train Model

python app/training/train.py

Build FAISS Indexes

python app/training/build_faiss_index.py

Run FastAPI Server

uvicorn app.main:app --reload

Swagger Documentation

http://127.0.0.1:8000/docs

---

Project Highlights

• End-to-End ML Pipeline
• Intent Classification System
• Hybrid Search Architecture
• Multilingual Query Processing
• FastAPI REST API
• Feedback Analytics
• FAISS Vector Search
• BM25 Retrieval
• Modular Architecture

---

Future Improvements

• Docker Deployment
• CI/CD Pipeline
• MLflow Integration
• PostgreSQL Storage
• Automated Retraining
• Monitoring & Observability
• Model Registry
• Cloud Deployment

---

Author

Kevin

Machine Learning Engineer | MLOps Enthusiast