Delphi - AI-Powered Drive

Delphi is a modern, intelligent file management system that combines the familiarity of a traditional cloud drive with powerful AI capabilities. It features instant search, folder management, seamless document previews, and intelligent reminder extraction.

🚀 Features

• Smart Storage: Upload, organize, and manage files and folders.
• AI Chat Assistant: Interactive RAG-powered chatbot to query your documents and get instant answers.
• Intelligent Reminders: Automatically extracts actionable items, deadlines, and due dates from uploaded documents (bills, appointments, contracts, etc.).
• Overdue Tracking: Automatically detects and marks overdue reminders with clear indicators.
• Hybrid Search: Powered by Meilisearch with Together AI embeddings for lightning-fast semantic and keyword search.
• Document Analysis: AI-powered extraction of summaries, keywords, and actionable items from PDFs and images.
• Document Preview: Built-in preview for PDFs and images without downloading, with keyboard navigation.
• Contextual Actions: Move, delete, and manage files with ease.
• Modern UI: A clean, responsive interface built with Shadcn UI and Tailwind CSS.

🛠 Tech Stack

• Frontend: Next.js 14 (App Router), React, TypeScript, Tailwind CSS, Shadcn UI, TanStack Query, Framer Motion.
• Backend: FastAPI (Python), SQLite, pyagentspec, wayflowcore.
• AI/LLM: Together AI (Qwen/Qwen2.5-VL-72B-Instruct for vision, openai/gpt-oss-120b for text analysis).
• Vector Database: ChromaDB for document embeddings.
• Search Engine: Meilisearch with Together AI embeddings for hybrid search.
• PDF Processing: PyMuPDF for text extraction from PDFs.
• Infrastructure: Local Meilisearch and ChromaDB instances.

📋 Prerequisites

Before you begin, ensure you have the following installed:

• Node.js (v18 or higher)
• Python (v3.8 or higher)
• Meilisearch (v1.13 or higher)

⚡️ Quick Start

1. Clone the Repository

git clone <repository-url>
cd delphi

2. Start Meilisearch (Search Engine)

Run Meilisearch locally. This is required for search functionality.

# Run Meilisearch (assuming it's installed via Homebrew or available in PATH)
meilisearch --env development --db-path ./meili_data

Keep this terminal window open.

3. Setup & Run Backend

Open a new terminal window:

cd backend

# Create virtual environment
python -m venv venv

# Activate virtual environment
# On macOS/Linux:
source venv/bin/activate
# On Windows:
# venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Start the server
cd .. && uvicorn backend.main:app --reload --port 8000

The backend will start at http://localhost:8000.

4. Environment Configuration (Required for AI Features)

To enable AI features (chat, document analysis, reminders, semantic search), you need to provide a Together AI API key.

1. Create a api.env file in the backend/ directory.
2. Add your Together AI API key:

TOGETHER_API_KEY=your-together-api-key-here

This key enables:

• AI Chat Assistant: RAG-powered chatbot for querying documents
• Document Analysis: Automatic extraction of summaries and keywords
• Reminder Extraction: Intelligent detection of actionable items and deadlines
• Semantic Search: Hybrid search with Together AI embeddings via Meilisearch

Get your API key at together.ai

5. Setup & Run Frontend

Open a third terminal window:

# Install dependencies
npm install

# Start the development server
npm run dev

The application will be available at http://localhost:3000.

🌱 Seeding Example Data

To populate the drive with realistic sample documents (Invoices, Receipts, Contracts, etc.) to test the search functionality:

1. Ensure the backend and Meilisearch are running.
2. Run the seed script:

cd backend
# Ensure venv is activated
python seed_database.py

This will create 10 sample documents with rich content that you can search for immediately.

🔍 Search Configuration

The search system uses Meilisearch with Together AI embeddings for hybrid semantic + keyword search.

• Endpoint: http://localhost:7700
• Index Name: documents
• Searchable Attributes: filename, content, summary, keywords
• Filterable Attributes: type, parent_folder_id
• Embeddings: Together AI vectors stored in Meilisearch for semantic similarity

The system automatically enables hybrid search when a TOGETHER_API_KEY is provided in api.env.

🤖 AI Features

Chat Assistant

• RAG-Powered: Queries your document collection using vector similarity
• Context-Aware: Provides answers with citations to source documents
• Real-time: Interactive chat widget with document preview integration

Reminder System

• Automatic Extraction: Analyzes uploaded documents for actionable items
• Smart Detection: Identifies bills, appointments, renewals, deadlines
• Categories: Automatically categorizes reminders (payment, appointment, subscription, etc.)
• Overdue Tracking: Marks past-due reminders with [OVERDUE] prefix
• Due Date Parsing: Extracts and tracks due dates from document content

Document Analysis

• Dual Processing: Separate workflows for images (OCR with vision model) and PDFs (text extraction)
• Summary Generation: AI-generated summaries for quick document understanding
• Keyword Extraction: Automatic tagging for improved searchability
• Vision Model: Qwen/Qwen2.5-VL-72B-Instruct for image-based documents
• Text Model: openai/gpt-oss-120b for PDF analysis and chat

🔄 Agent Workflows

Delphi uses pyagentspec and wayflowcore to orchestrate intelligent document processing through two specialized agent workflows. Each workflow is a directed graph of nodes that process documents sequentially.

Image Processing Workflow

Flow: Start → Vision Extraction → Summarization → Reminder Analysis → End

1. Start Node

   • Inputs: image_path, filename
   • Initializes the workflow with the path to an image document

2. Vision Extraction Node (analyze_image)

   • Tool: Vision LLM (Qwen/Qwen2.5-VL-72B-Instruct)
   • Purpose: Performs OCR on images to extract text content
   • Process:
     • Loads image bytes and determines format (JPEG, PNG, etc.)
     • Creates a prompt with system instructions and the image
     • Calls the vision model via Together AI API
     • Returns extracted text or empty string if no text found
   • Output: extracted_text

3. Summarization Node (summarize_and_extract_keywords)

   • Tool: Text LLM (openai/gpt-oss-120b)
   • Purpose: Generates document summary and extracts searchable keywords
   • Process:
     • Takes extracted text (skips if less than 10 characters)
     • Sends to text model with summarization prompt
     • Parses JSON response containing summary and keyword list
     • Handles markdown code blocks in LLM responses
   • Output: analysis_result (JSON with summary and keywords)

4. Reminder Analysis Node (analyze_for_reminder)

   • Tool: Text LLM (openai/gpt-oss-120b)
   • Purpose: Detects actionable items and deadlines
   • Process:
     • Combines extracted text, summary, and keywords into context
     • Sends to text model with strict reminder analysis prompt
     • Identifies if document requires action (bills, appointments, renewals, deadlines)
     • Extracts due dates, categories, action titles, and descriptions
     • Returns structured JSON with reminder metadata
   • Output: reminder_data (JSON with requires_action, category, due_date, action_title, action_description)

5. End Node

   • Aggregates all outputs: text, analysis, reminder
   • Returns complete document processing results

Data Flow:

• image_path → Vision Extraction
• extracted_text → Summarization + Reminder Analysis
• filename → Reminder Analysis (for context)
• analysis_result → Reminder Analysis (provides keywords/summary context)
• All outputs → End Node

PDF Processing Workflow

Flow: Start → Text Extraction → Summarization → Reminder Analysis → End

1. Start Node

   • Inputs: pdf_path, filename
   • Initializes the workflow with the path to a PDF document

2. PDF Text Extraction Node (extract_pdf_text)

   • Tool: PyMuPDF (fitz library)
   • Purpose: Extracts raw text from PDF files without requiring vision model
   • Process:
     • Opens PDF using PyMuPDF
     • Iterates through all pages
     • Extracts text from each page using .get_text()
     • Concatenates all page text
     • Validates minimum text length (10 characters)
   • Output: extracted_text
   • Note: Much faster than vision model since it reads native PDF text

3. Summarization Node (summarize_and_extract_keywords)

   • Identical to image workflow
   • Processes extracted PDF text to generate summary and keywords

4. Reminder Analysis Node (analyze_for_reminder)

   • Identical to image workflow
   • Analyzes PDF content for actionable items and deadlines

5. End Node

   • Aggregates all outputs: text, analysis, reminder
   • Returns complete document processing results

Data Flow:

• pdf_path → PDF Text Extraction
• extracted_text → Summarization + Reminder Analysis
• filename → Reminder Analysis (for context)
• analysis_result → Reminder Analysis (provides keywords/summary context)
• All outputs → End Node

Key Differences Between Workflows

Aspect	Image Workflow	PDF Workflow
Extraction Method	Vision LLM (Qwen) via API	PyMuPDF text extraction
Speed	Slower (API call + model inference)	Faster (native text reading)
Accuracy	Good for scanned docs, photos	Perfect for digital PDFs
Cost	API credits per image	Local, no API cost
Use Case	Screenshots, scanned docs, photos	Native PDFs with selectable text

Workflow Orchestration

Both workflows use:

• Control Flow Edges: Define sequential execution order between nodes
• Data Flow Edges: Pass outputs from one node as inputs to another
• Tool Registry: Maps tool names to Python function implementations
• AgentSpecLoader: Loads and executes the flow with registered tools
• Conversation Context: Maintains state throughout the workflow execution

The workflows automatically handle errors, parse JSON responses, and provide fallback values if any step fails.

📂 Project Structure

.
├── app/                    # Next.js App Router pages
│   ├── page.tsx            # Main drive page with document browsing
│   ├── reminders/          # Reminders page with overdue tracking
│   └── layout.tsx          # Root layout with sidebar and chat widget
├── backend/                # FastAPI Backend
│   ├── main.py             # API Endpoints & Document Processing Pipeline
│   ├── store.py            # Database & File System Logic
│   ├── meilisearch_client.py # Search Engine Client with Together AI
│   ├── embeddings.py       # Vector embedding generation
│   ├── extractor.py        # Document text extraction & AI analysis
│   ├── chat_agent.py       # RAG-powered chat agent
│   ├── prompt.py           # LLM prompts for analysis & reminders
│   ├── seed_database.py    # Data Seeding Script
│   ├── metadata.db         # SQLite Database (documents, folders, reminders)
│   ├── uploads/            # File Storage
│   └── chroma_db/          # ChromaDB vector storage
├── components/             # React Components
│   ├── chat/               # Chat widget with RAG integration
│   ├── drive/              # File Grid, List, Folder components
│   ├── preview/            # Document Preview Modal with keyboard nav
│   ├── search/             # Search Bar & Results
│   ├── layout/             # Sidebar & Header
│   └── ui/                 # Shadcn UI primitives
├── hooks/                  # Custom React Hooks
│   ├── useChat.ts          # Chat assistant integration
│   ├── useDocuments.ts     # Document CRUD operations
│   ├── useSearch.ts        # Search functionality
│   └── useUpload.ts        # File upload with progress
├── contexts/               # React Context providers
├── meili_data/             # Meilisearch data directory
└── public/                 # Static assets