Indexify

A powerful semantic search engine that combines traditional text search with vector similarity using Elasticsearch and machine learning embeddings.

video1.mp4

Overview

Indexify is a comprehensive search solution that leverages Google Custom Search API, Elasticsearch, and transformer-based embeddings to provide intelligent search capabilities. It combines traditional keyword search with semantic understanding to deliver more relevant results.

✨ Features

• 🔍 Hybrid Search: Combining text and vector similarity
• 🤖 ML-Powered: Text embeddings using sentence-transformers
• 📊 Analytics: Search statistics and trending queries tracking
• 💡 Smart Suggestions: Based on user behavior
• 🔄 Auto-Fetch: Content from Google Custom Search when needed
• 🎯 Advanced Search: Multiple criteria (title, author, date, keywords)

Built with

• Next.js
• Tailwind CSS
• Elastic Search
• FastAPI
• Google Custom Search

🏗️ Architecture

graph TD
    subgraph Frontend
        UI[User Interface]
        SearchBar[Search Bar]
        Results[Results Display]
        Suggestions[Search Suggestions]
    end

    subgraph Backend API
        API[FastAPI Backend]
        SearchHandler[Search Handler]
        SuggestionEngine[Suggestion Engine]
        IndexManager[Index Manager]
    end

    subgraph External Services
        Google[Google Custom Search API]
        Transform[Sentence Transformer]
    end

    subgraph Elasticsearch
        ESIndex[Search Index]
        Stats[Search Statistics]
        Vectors[Vector Storage]
    end

    UI --> |Search Query| SearchBar
    SearchBar --> |API Request| API
    SearchBar --> |Get Suggestions| SuggestionEngine

    API --> |Process Query| SearchHandler
    SearchHandler --> |Vector Search| ESIndex
    SearchHandler --> |Update Stats| Stats

    API --> |New Content Request| Google
    Google --> |Raw Results| IndexManager
    IndexManager --> |Generate Embeddings| Transform
    Transform --> |Store Vectors| Vectors

    SuggestionEngine --> |Fetch Trends| Stats
    ESIndex --> |Return Results| Results
    Stats --> |Popular Searches| Suggestions

    style UI fill:#f9f,stroke:#333,color:#000
    style ESIndex fill:#69b,stroke:#333,color:#000
    style Transform fill:#9cf,stroke:#333,color:#000
    style Google fill:#4a8,stroke:#333,color:#000

Loading

Backend Structure

backend/
├── app/
│   ├── api/routes/
│   │   └── routes.py
│   ├── core/
│   │   ├── models.py
│   │   ├── elasticsearch.py
│   │   ├── suggestion.py
│   │   ├── custom_search.py
│   │   ├── utils.py
│   │   ├── client.py
│   │   ├── index.py
│   │   └── documents.py
│   └── main.py
├── .env
├── requirements.txt
└── README.md

Frontend Structure

frontend/indexify/
├── .next/
├── node_modules/
├── public/
├── src/
│   └── app/
│       ├── globals.css
│       ├── layout.tsx
│       └── page.tsx
├── components/
├── └── SearchBar.tsx
├── config/
│   └── constants.ts
├── hooks/
│   ├── useSearch.ts
│   └── useSuggestions.ts
├── types/
│   └── index.ts
├── .env
├── .gitignore
├── eslint.config.mjs
├── next-env.d.ts
├── next.config.ts
├── package-lock.json
├── package.json
├── postcss.config.mjs
├── README.md
├── tailwind.config.ts
└── tsconfig.json

Technical Details

Elasticsearch Index Mapping

The system uses a sophisticated index mapping with the following fields:

"mappings": {
                "properties": {
                    "title": {
                        "type": "text",
                        "analyzer": "custom_text_analyzer",
                        "fields": {
                            "keyword": {"type": "keyword"},
                            "completion": {
                                "type": "completion",
                                "analyzer": "custom_text_analyzer"
                            }
                        }
                    },
                    "author": {"type": "keyword"},
                    "publication_date": {"type": "date"},
                    "abstract": {"type": "text", "analyzer": "custom_text_analyzer"},
                    "keywords": {
                        "type": "keyword",
                        "fields": {
                            "text": {
                                "type": "text",
                                "analyzer": "custom_text_analyzer"
                            }
                        }
                    },
                    "content": {"type": "text", "analyzer": "custom_text_analyzer"},
                    "vector": {"type": "dense_vector", "dims": vector_dims},
                    "search_count": {"type": "long"}
                }
            }

• title: Text field with keyword and completion sub-fields
• author: Keyword field for exact matching
• publication_date: Date field
• abstract: Text field with custom analyzer
• keywords: Keyword field with text sub-field
• content: Text field with custom analyzer
• vector: Dense vector field for semantic search
• search_count: Long field for tracking popularity

Embedding Process

Indexify uses the sentence-transformers/all-MiniLM-L6-v2 model to generate semantic text embeddings that capture the meaning of content. Here's how the process works:

1. Model Initialization

model_name = "sentence-transformers/all-MiniLM-L6-v2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name)

2. Content Processing Flow

graph TD
    A[Input Text] -->|Tokenization| B[Tokens]
    B -->|Model Processing| C[Raw Embeddings]
    C -->|Extract CLS Token| D[Final Vector]
    D -->|Store| E[Elasticsearch]

Loading

3. Technical Implementation

Text Preprocessing & Embedding Generation

def generate_embedding(text: str) -> list[float]:
    # Tokenize with truncation for long texts
    inputs = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        max_length=512
    )

    # Generate embeddings without gradient computation
    with torch.no_grad():
        outputs = model(**inputs)

    # Extract CLS token embedding
    embedding = outputs.last_hidden_state[:, 0, :].squeeze().tolist()
    return embedding

Search Process with Embeddings

def vector_text_search(client, index_name, query_text, query_vector):
    query = {
        "query": {
            "script_score": {
                "query": {
                    "multi_match": {
                        "query": query_text,
                        "fields": ["title^3", "abstract^2", "content"]
                    }
                },
                "script": {
                    "source": "cosineSimilarity(params.query_vector, 'vector') + 1.0",
                    "params": {"query_vector": query_vector}
                }
            }
        }
    }

4. Processing Pipeline

1. Input Processing:

   • Combines title and snippet for search results
   • Truncates to 512 tokens maximum
   • Handles special tokens automatically

2. Vector Generation:

   • Converts tokens to model inputs
   • Processes through transformer model
   • Extracts CLS token representation
   • Converts to float list format

3. Search Integration:

   • Stores vectors in Elasticsearch
   • Uses cosine similarity for matching
   • Combines with text-based relevance
   • Boosts results based on field importance

4. Result Scoring:

   • Base text similarity score
   • Vector similarity contribution
   • Optional keyword presence boost
   • Field-specific weight multipliers

The embedding system enhances search accuracy by:

• Capturing semantic relationships
• Understanding context beyond keywords
• Enabling similarity-based matching
• Supporting hybrid ranking strategies

Search Features

1. Vector Text Search

• Combines traditional text matching with vector similarity
• Uses script scoring for hybrid ranking
• Supports fuzzy matching and field boosting

2. Advanced Search

• Multi-criteria search (title, author, date range, keywords)
• Customizable result size
• Sort by relevance and date

3. Search Suggestions

• Based on previous searches and trending queries
• Tracks and updates search statistics
• Provides real-time completion suggestions

API Routes

Core Endpoints

POST /api/search
POST /api/advanced-search
GET /api/suggestions

Interested in Contributing?

If you're interested, please see Backend and Frontend Guidelines.

License

This project is licensed under the MIT License - see the LICENCE file for details.