https://github.com/mmonire/librechat

Simple RAG chatbot designed to answer questions about Linux, free software, and influential figures like Linus Torvalds and Richard Stallman.
https://github.com/mmonire/librechat

chatbot cohere rag

Last synced: 2 months ago
JSON representation

Simple RAG chatbot designed to answer questions about Linux, free software, and influential figures like Linus Torvalds and Richard Stallman.

• Host: GitHub
• URL: https://github.com/mmonire/librechat
• Owner: Mmonire
• Created: 2025-07-10T09:18:56.000Z (12 months ago)
• Default Branch: main
• Last Pushed: 2025-07-10T09:38:44.000Z (12 months ago)
• Last Synced: 2025-07-10T17:40:55.886Z (12 months ago)
• Topics: chatbot, cohere, rag
• Language: Jupyter Notebook
• Homepage:
• Size: 20.5 KB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.markdown

Awesome Lists containing this project

README

          
# 🤖 LibreChat: Linux Philosophy Chatbot

## 📖 Overview

**LibreChat** is a Retrieval-Augmented Generation (RAG) chatbot designed to answer questions about Linux, free software, and influential figures like Linus Torvalds and Richard Stallman. It leverages diverse data sources, including a PDF book, web content, Wikipedia pages, and HTML files from stallman.org, to provide concise, accurate responses in Persian (up to 4 words). Built using **LangChain** and **Cohere**, LibreChat processes documents, embeds them in a vector store, and retrieves relevant information to answer 16 predefined questions for automated evaluation.

## 🎯 Objectives

- **Build a RAG Chatbot**: Implement a retrieval-augmented system to answer questions based on provided data sources.

- **Process Diverse Data**: Load and preprocess PDF, web, Wikipedia, and HTML content into unified documents.

- **Ensure Concise Responses**: Generate short (≤4 words) Persian answers in a structured JSON format.

- **Achieve High Accuracy**: Correctly answer at least 12 out of 16 evaluation questions to meet project requirements.

## ✨ Features

- **Multi-Source Data Handling**:

  - 📄 **PDF**: Extracts text from "Just for Fun" (Persian translation) using `PyPDFium2Loader`.

  - 🌐 **Web**: Scrapes "Linux and Life" from linuxbook.ir using `WebBaseLoader`.

  - 📚 **Wikipedia**: Loads Persian pages on Linux, free software, and key figures with `WikipediaLoader`.

  - 🖥️ **HTML**: Processes stallman.org pages using `DirectoryLoader` and `BSHTMLLoader`.

- **RAG Architecture**:

  - Splits documents into chunks (`RecursiveCharacterTextSplitter`).

  - Embeds using Cohere’s multilingual model (`embed-multilingual-light-v3.0`).

  - Stores embeddings in a Chroma vector store for semantic search.

  - Uses Cohere reranking (`rerank-multilingual-v3.0`) for improved retrieval.

- **Structured Output**: Returns JSON dictionaries with `question_number` (int) and `answer` (str) for evaluation.

- **Few-Shot Prompting**: Guides the model with example questions and answers to ensure format compliance.

## 🛠 Prerequisites

To run this project, you need:

- **Python** 3.9 or higher.

- **Libraries**:

  - `langchain`, `langchain-community`, `langchain-cohere`, `langchain-chroma`

  - `pypdfium2`, `pandas`, `beautifulsoup4`

- **Cohere API Key**: Obtain from [Cohere](https://cohere.ai/) and set as `COHERE_API_KEY`.

- **Dataset Files** (in `data` folder):

  - `justforfun_persian.pdf`

  - `html/` (containing stallman.org HTML files)

- **Environment**: Google Colab with GPU recommended for Hugging Face models (optional).

## 📦 Installation

1. Clone or download the repository:

   ```bash

   git clone 

   ```

2. Navigate to the project directory:

   ```bash

   cd 

   ```

3. Install required libraries:

   ```bash

   pip install langchain langchain-community langchain-cohere langchain-chroma pypdfium2 pandas beautifulsoup4

   ```

4. Set up the Cohere API key:

   ```python

   import os

   os.environ["COHERE_API_KEY"] = "your-api-key"

   ```

5. Ensure dataset files are in the `data` folder.

## 🚀 Usage

1. Open `librechat.ipynb` in Jupyter Notebook or Google Colab.

2. Run cells sequentially to:

   - Load and preprocess data (PDF, web, Wikipedia, HTML).

   - Split documents into chunks and embed them in a Chroma vector store.

   - Set up the RAG chain with Cohere LLM and reranker.

   - Answer 16 evaluation questions and save responses.

3. Generate the submission file:

   ```bash

   python -m notebook librechat.ipynb

   ```

4. Output:

   - `result.zip` containing:

     - `librechat.ipynb`

     - `answers.json` (16 JSON dictionaries with question numbers and answers)

## 📊 Code Structure

- **Data Loading**:

  - Loads PDF (`PyPDFium2Loader`), web (`WebBaseLoader`), Wikipedia (`WikipediaLoader`), and HTML (`DirectoryLoader` with `BSHTMLLoader`).

  - Saves preprocessed documents to `processed_docs.json`.

- **Document Splitting**:

  - Uses `RecursiveCharacterTextSplitter` (chunk size: 500, overlap: 50).

  - Saves chunks to `splitted_docs.json`.

- **Vector Store**:

  - Embeds documents with `CohereEmbeddings` (`embed-multilingual-light-v3.0`).

  - Stores in Chroma (`persist_directory="./vec_store_chroma"`).

- **RAG Chain**:

  - Retrieves documents with `vector_store.as_retriever` (k=50).

  - Reranks with `CohereRerank` (top_n=20).

  - Generates answers using `ChatCohere` with a custom prompt template.

  - Outputs JSON dictionaries via `rag_chain` function.

- **Submission**:

  - Combines answers (`answer1` to `answer16`) into `answers.json`.

  - Zips with notebook into `result.zip`.

## 🔍 Evaluation

- **Questions**: 16 predefined questions (e.g., "پرسش ۴: چه کسی بنیاد نرم‌افزارهای آزاد را بنا نهاد؟").

- **Output Format**: Each answer is a JSON dictionary:

  ```json

  {"question_number": 4, "answer": "ریچارد استالمن"}

  ```

- **Constraints**:

  - Answers must be ≤4 words in Persian.

  - At least 12 correct answers required for success.

## 🎓 Credits

Developed as part of a Quera academic project, leveraging LangChain and Cohere for RAG-based chatbot implementation.