# GEMINI.md - Project Context for Gemini CLI

This document provides a comprehensive overview of the "Rapport-automatique" project for the Gemini CLI, enabling it to understand the project's purpose, architecture, and key components for effective collaboration.

## Project Overview

The goal of this project is to create an AI agent that can automatically write an internship report. The agent uses a Retrieval-Augmented Generation (RAG) system to source information from a collection of notes (weekly internship reports) provided in the `documents_projet/` directory.

The project is built in Python and leverages the LangChain and LangGraph frameworks to create a sophisticated agent.

**Key Technologies:**

*   **Orchestration:** LangChain & LangGraph
*   **LLM:** `mistral-large-latest` via `ChatMistralAI`
*   **Vector Database (RAG):** ChromaDB (persisted in `chroma_db/`)
*   **Embeddings:** `jinaai/jina-embeddings-v3` from HuggingFace
*   **Document Loading:** `Unstructured` (for `.txt` files)
*   **Web Search:** Tavily
*   **Experiment Tracking:** MLflow

## Architecture

The system is designed as a LangGraph agent with a clear, cyclical flow:

1.  **Start (LLM Call):** The agent starts by calling the Mistral LLM (`reponse_question` node) with the current conversation history.
2.  **Tool Decision:** The LLM decides whether to generate a direct response or use one of its available tools.
3.  **Conditional Routing:** The `should_continue` function checks the LLM's output. If tool calls are present, the graph transitions to the `tool_node`. Otherwise, the session ends.
4.  **Tool Execution:** The `tool_node` executes the requested tools (e.g., `search_in_files` for RAG, `internet_search`, file I/O).
5.  **Loop:** The output of the tools is passed back to the LLM (`reponse_question` node) for it to process the results and decide the next action, continuing the cycle.

The agent's state (`CustomState`) is explicitly managed and includes conversation history, a `todo` list for task management, and the query and results from the RAG system.

## Building and Running

### 1. Setup

**a. Install Dependencies:**
First, set up and activate a Python virtual environment. Then, install the required packages.

```bash
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
```

**b. Configure Environment Variables:**
Copy the template `.env.template` file and fill in your API keys (e.g., for Mistral, Tavily).

```bash
cp AgentReact/.env.template AgentReact/.env
# Edit AgentReact/.env with your credentials
```

### 2. Data Ingestion (RAG Setup)

Place your source documents (as `.txt` files) into the `documents_projet/` directory at the project root. Then, run the initialization script to populate the Chroma vector database.

```bash
python RAG/init.py
```

### 3. Running the Agent

The main entry point for the agent is `AgentReact/start.py`.

```bash
python AgentReact/start.py
```

This script will invoke the agent graph with a hardcoded sample question and print the resulting messages.

## Development Conventions

*   **Modularity:** The code is well-structured into directories for the agent (`AgentReact`), RAG components (`RAG`), and data (`documents_projet`). The agent's logic is further divided into `agent.py` (graph), `nodes.py`, `state.py`, and `tools.py`.
*   **Singleton Pattern:** The `VectorDatabase` is implemented as a Singleton to ensure a single, shared instance throughout the application.
*   **State Management:** The agent's state is explicitly defined in `AgentReact/utils/state.py`, making it clear what information is tracked across turns.
*   **Human-in-the-Loop:** The `ask_human` tool provides a mechanism for the agent to request user input, although the full "supervised tools" workflow from the diagram is not yet implemented.
*   **Roadmap:** The `roadmap.md` file tracks the project's progress and outlines future development goals, such as moving from ChromaDB to PG Vector and adding PDF generation.