chatBOTr - Open Source R Shiny Chatbot

A Retrieval-Augmented Generation (RAG) chatbot built with R Shiny. This application is a boilerplate implementation that may serve as an intelligent search assistant for specific websites (configurable via environment variables). It performs a Google Custom Search, downloads the results (HTML/PDF), feeds them into a Large Language Model (LLM), and provides strict, fact-based answers with citations. Alternatively, a variant that uses Google Gemini directly is also implemented.

Key Features

• RAG Pipeline: Automatically searches, downloads, and processes context from a target website.
• Strict Grounding: The system prompt is engineered to answer only using the provided documents, minimizing hallucinations.
• Bilingual Support: Full support for English and German. The UI, system prompts, and LLM answers adapt to the selected language.
• Multi-Provider Support:
  • LLM: Compatible with any OpenAI-API compliant endpoint (e.g., Local Llama, vLLM, SSPCloud).
  • Google Gemini: Native support via the ellmer package.
• Dynamic Geo-Restriction: Automatically detects the target country based on the website TLD (e.g., .at -> Austria) to optimize search queries.

Prerequisites

The application relies on environment variables for API keys and configuration. You can create an .Renviron file or define the variables directly.

1. Google Custom Search (Required)

You need a Google Programmable Search Engine (CSE) to fetch results.

• API Key: Get it from Google Cloud Console.
• CX ID: The Search Engine ID.
• Site: The domain you want the bot to search (e.g., www.statistik.at).

2. LLM Provider

• Custom / Local: An endpoint running an OpenAI-compatible API.
• Google Gemini: Requires the standard Google API Key.

---

Configuration (.Renviron)

Create a file named .Renviron in the root of your project:

# --- LLM Endpoint (OpenAI Compatible) ---
# Used if provider is set to "Custom / Local"
LLM_BASE_URL="http://your-local-llm:8080/api/"
LLM_TOKEN="your-access-token"

# --- Google Services ---
# Required for Search AND for Gemini Model
GCS_APIKEY="your-google-api-key"
GCS_CX="your-custom-search-engine-id"

# --- App Configuration ---
# The website to search and analyze. 
# The bot derives geo-restrictions from the TLD (e.g., .at = Austria, .de = Germany)
GCS_SITE="www.statistik.at"

# Cache file path (for poor mans caching)
LLM_CACHING_DB="file-upload-cache.rds"

Note that you can also set the Environment-Variables directly, e.g. when you want to dockerize this setup or deploy in via Posit Connect.

---

Installation & Usage

1. Open the project in RStudio.
2. Install dependencies:

   install.packages(c("shiny", "shinyalert", "cli", "httr", "jsonlite", "digest", "bslib", "fs", "glue", "markdown", "ellmer"))

3. Ensure .Renviron is present in the root.
4. Run the app:

   shiny::runApp()

---

How it Works

1. Input: User enters a query and selects a language (English/German).
2. Optimization: An LLM rewrites the query for Google Search. It automatically adds site-specific or geo-specific restrictions based on the GCS_SITE environment variable (e.g., adding (Austria OR AT) if the site is .at).
3. Search: Google Custom Search API returns top N results.
4. Ingestion: The app downloads the content (HTML text or PDFs) from those URLs.
5. Context Construction: A system prompt is generated containing:
   • Strict behavioral instructions in the selected language.
   • The content of the downloaded files.
   • Metadata for citations.
6. Generation: The LLM is instructed to generate an answer using only the provided context, citing sources like [1] {file.pdf}.