Projet_assistantRH_LLM

This project was developed in the context of an LLM training module, as a practical exercise in building an assistant tool for Human Resources (RH) using RAG (Retrieval-Augmented Generation) techniques.

It provides a pipeline to:

• Build a RAG from HR-related data,
• Query the assistant with natural language,
• Interact via a simple interface,
• Evaluate the quality of answers against a dataset.

---

🚀 Installation

The project uses uv for environment and dependency management.

1. Clone the repository and switch to the organized branch:

   git clone https://github.com/DidiCi/Projet_assistantRH_LLM.git
   cd Projet_assistantRH_LLM

2. Install the environment with uv:

   uv sync

3. Set up configuration:

   • Obtain a Google API key and save it in a .env file at the project root:

     GOOGLE_API_KEY=your_api_key_here
     

   • Input/output folders and other options can be configured in rag/config.py.

4. Prepare the data:

   • Place your CV files (the documents in PDF format to be analyzed) in:

     data/raw/
     

---

🧠 Usage

1. Build and test the RAG

Create the RAG:

uv run python rag/main.py

You can also ask a direct question when running it:

uv run python rag/main.py --question "Qui parle italien?"

2. Run the interface

Once the RAG is created, launch the Streamlit interface:

uv run streamlit run app/interface.py

This provides a user-friendly way to interact with the assistant.

---

📊 Evaluation

The project includes tools to evaluate the RAG’s answers.

1. Define your test set by editing:

   evaluation/evaluation_dataset.json
   

   Add questions and their expected answers.

2. Run the evaluation pipeline:

   uv run python evaluation/evaluation_llm.py
   uv run python evaluation/evaluation_score.py

This will generate scores and metrics about the assistant’s accuracy and relevance.

---

📂 Project Structure

Path	Description
app/	Streamlit interface for interacting with the assistant.
rag/	Core RAG implementation (retrieval, embeddings, pipeline).
rag/config.py	Configuration file for input/output folders and settings.
data/raw/	Folder where input CVs must be placed.
evaluation/	Scripts and datasets for evaluating RAG answers.
evaluation/evaluation_dataset.json	JSON dataset of questions & answers for evaluation.
.env	Must contain the Google API key.
pyproject.toml, uv.lock	Project dependencies managed by uv.

---

🔧 Requirements

• uv
• Python (version specified in .python-version)
• Google API key

Dependencies are automatically installed via uv sync.

---

📌 Context

This repository was created as part of an LLM formation module, to practice:

• Using RAG for domain-specific assistants,
• Managing configurations and pipelines,
• Evaluating model performance systematically,
• Building a minimal interactive application.

---