DocSense AI

DocSense AI is an advanced agentic system powered by LlamaIndex and Google Gemini (gemini-2.0-flash). It processes, indexes, and queries complex PDF documents using Retrieval-Augmented Generation (RAG) while persisting session records dynamically via local tool integration.

System Architecture

graph TD
    User([User CLI]) -->|Queries| Agent[ReAct Agent]
    Agent -->|1. Route Request| Router{Tool Router}
    Router -->|If PDF Query| RAG[LlamaIndex Query Engine]
    Router -->|If Note Saving| Notes[Note Saver Tool]
    RAG -->|Retrieve Context| Storage[(Vector Store Index)]
    Storage -.->|Token Chunking| PDF[Sri_Lanka.pdf]
    RAG -->|Generate Answer| LLM[Google Gemini 2.0]
    Notes -->|Write Notes| File[data/notes.txt]
    LLM -->|Answer| Agent
    Agent -->|Final Output| User

    %% Styles with high contrast dark text for dark-mode safety
    style User fill:#e1f5fe,stroke:#01579b,stroke-width:2px,color:#0a2540
    style Agent fill:#ede7f6,stroke:#4a148c,stroke-width:2px,color:#0a2540
    style Router fill:#fff3e0,stroke:#e65100,stroke-width:2px,color:#0a2540
    style RAG fill:#e8f5e9,stroke:#1b5e20,stroke-width:2px,color:#0a2540
    style Notes fill:#fbe9e7,stroke:#bf360c,stroke-width:2px,color:#0a2540
    style Storage fill:#e8eaf6,stroke:#1a237e,stroke-width:2px,color:#0a2540
    style PDF fill:#fce4ec,stroke:#880e4f,stroke-width:2px,color:#0a2540
    style LLM fill:#e0f2f1,stroke:#004d40,stroke-width:2px,color:#0a2540
    style File fill:#efebe9,stroke:#3e2723,stroke-width:2px,color:#0a2540

Loading

Core Features

• Intelligent ReAct Agent: Autonomously determines whether to consult knowledge bases or execute procedural tools (e.g. saving summaries).
• API Rate-Limit Optimizer: Special token-splitting batch loader with custom delay buffers to respect free-tier API quotas (bypasses Gemini 429 Resource Exhausted errors).
• Persistent Vector Indexing: Caches document vectors locally to guarantee instantaneous search on subsequent runs.

Technical Details

• Knowledge Source: The RAG context is populated from a comprehensive PDF handbook containing geographic, cultural, and political profiles of Sri Lanka.
• Local Notes Persistence: Integrates a note_saver FunctionTool to record session logs directly into data/notes.txt on command.
• Adaptive API Throttling: Embeds nodes using the new gemini-embedding-2 model in 5-node batches, with 3-second intervals and an automatic 60-second recovery timeout block. This prevents API lockouts on Google's free tier (15 requests per minute limit).
• Model Configurations:
  • LLM: gemini-2.0-flash (via google-genai unified SDK)
  • Embeddings: gemini-embedding-2

---

Getting Started

Prerequisites

Ensure you have Python 3.10+ installed.

Installation & Setup

1. Clone the repository:

   git clone https://github.com/SihanUdayaratna03/DocSense-AI.git
   cd DocSense-AI

2. Install dependencies:

   pip install -r requirements.txt

3. Configure Environment: Create a .env file in the root folder:

   GEMINI_API_KEY=your_gemini_api_key_here

4. Launch Agent:

   python main.py