Sherwinj10/ai-knowledge-base

AI Knowledge Base Agent

An intelligent document assistant capable of ingesting, retrieving, and answering questions from PDF and Text documents using Retrieval-Augmented Generation (RAG) with session memory.

Overview

The AI Knowledge Base Agent is designed to bridge the gap between static documents and dynamic information retrieval. By leveraging RAG, it allows users to upload PDF or Text documents and ask natural language questions about them.

Unlike generic AI models, this agent grounds its answers strictly in the provided content, ensuring accuracy and relevance for specific use cases like HR policies, technical manuals, or legal contracts.

Features & Limitations

Features

• —Document Ingestion: Supports PDF and TXT files using LangChain's PyPDFLoader and TextLoader.
• —Context-Aware Answers: Uses RAG to retrieve relevant document chunks for accurate answering.
• —Structured Output: Answers are formatted with bullet points, a summary, and confidence levels.
• —Source Citations: Provides transparency by citing the specific document chunks used.
• —Local Privacy: Uses local HuggingFace embeddings (all-MiniLM-L6-v2) to generate vectors on your machine.
• —Modern UI: Built with a clean HTML/JS Frontend and a fast FastAPI Backend.
• —Cost-Effective: Uses Google Gemini 2.5 Flash for high-speed, low-cost inference.

Limitations

• —Text-Only Analysis: Currently extracts text only; does not process images or tables within PDFs (no OCR).
• —Single File Upload: Optimized for processing one file at a time.
• —Session Memory: Chat history is session-based and clears when you refresh the page.

Tech Stack & APIs Used

• —Frontend: HTML, CSS, JavaScript (Vanilla)
• —Backend: FastAPI (Python)
• —Ingestion: LangChain (PyPDFLoader, TextLoader)
• —LLM: Google Gemini 2.5 Flash (via langchain-google-genai)
• —Embeddings: HuggingFace all-MiniLM-L6-v2 (Local execution)
• —Vector Database: ChromaDB (Local persistence)

Setup & Run Instructions

Prerequisites

• —Python 3.10 or higher
• —A Google Cloud API Key (for Gemini)

Installation

1. 1.Clone the Repository

    git clone <your-repo-url>
    cd AI_Knowledge_Base_Agent

1. 1.Create a Virtual Environment

    python3 -m venv venv
    source venv/bin/activate  # On Windows: venv\Scripts\activate

1. 1.Install Dependencies

    pip install -r backend/requirements.txt

1. 1.Set Up Environment Variables Create a .env file in backend/.env and add your API key:

    GOOGLE_API_KEY=your_google_key

Running the App

1. 1.Start the Backend Server:

    uvicorn backend.main:app --reload

1. 1.Access the App: Open your browser and navigate to http://localhost:8000.

Potential Improvements

• —Multi-File Support: Allow uploading entire folders or multiple PDFs at once.
• —OCR Integration: Add support for scanned documents and images using Tesseract or specialized models.
• —Chat History Persistence: Save chat history to a database (SQLite/Postgres).
• —Docker Support: Containerize the application for easier deployment.

Architecture Diagram

graph TD

    subgraph Ingestion
        A[PDF TXT Document] -->|PyPDFLoader TextLoader| B(Raw Text)
        B -->|RecursiveSplitter| C[Text Chunks]
    end

    subgraph Embedding_Storage
        C -->|HuggingFace Embeddings| D[(ChromaDB)]
    end

    subgraph Retrieval_Generation
        E[User Query] -->|Embed| F[Query Embedding]
        F -->|Vector Search| D
        D -->|Retrieve Top k| G[Relevant Context]
        G -->|Context plus Query| H[Gemini 25 Flash]
        H -->|Generate| I[Structured Answer]
    end
    
    subgraph UI
        J[Frontend HTML JS] -->|API| K[Backend FastAPI]
        K --> J
        K --> H
        H --> K
    end