docs: Reorganize documentation and rewrite README for Library RAG

Major documentation cleanup and restructuring:

1. Documentation reorganization:
   - Created docs/migration-gpu/ directory
   - Moved 6 migration-related MD files to docs/migration-gpu/
   - Moved project_progress.md to docs/

2. Complete README.md rewrite:
   - Comprehensive explanation of dual RAG system
   - Clear documentation of 5 Weaviate collections:
     * Library Philosophique: Work, Chunk_v2, Summary_v2
     * Memory Ikario: Thought, Conversation
   - GPU embedder architecture (BAAI/bge-m3, RTX 4070, 1024-dim)
   - Quick start guide with installation steps
   - Usage examples for all features (search, chat, memories, upload)
   - Performance metrics (30-70x faster ingestion)
   - Troubleshooting section
   - Project structure overview

3. Benefits:
   - Reduced root-level clutter (7 MD files → organized structure)
   - Clear separation: migration docs vs project docs
   - User-friendly README focused on usage, not implementation
   - Easier navigation for new users

Files moved:
- BUG_REPORT_WEAVIATE_CONNECTION.md → docs/migration-gpu/
- DIAGNOSTIC_ARCHITECTURE_EMBEDDINGS.md → docs/migration-gpu/
- MIGRATION_GPU_EMBEDDER_SUCCESS.md → docs/migration-gpu/
- TEST_CHAT_GPU_EMBEDDER.md → docs/migration-gpu/
- TEST_FINAL_GPU_EMBEDDER.md → docs/migration-gpu/
- TESTS_COMPLETS_GPU_EMBEDDER.md → docs/migration-gpu/
- project_progress.md → docs/

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
2026-01-09 12:49:42 +01:00
parent eb2bf45281
commit 7cbcdeb476
8 changed files with 384 additions and 475 deletions

857
README.md
View File

@@ -1,554 +1,463 @@
# Autonomous Coding Agent Demo (Linear-Integrated) # Library RAG - Système de Recherche Philosophique Avancé
A minimal harness demonstrating long-running autonomous coding with the Claude Agent SDK. This demo implements a two-agent pattern (initializer + coding agent) with **Linear as the core project management system** for tracking all work. Système RAG (Retrieval-Augmented Generation) dual pour la recherche philosophique et la mémoire conversationnelle, propulsé par GPU embedder et Weaviate.
## Key Features ## 🎯 Vue d'Ensemble
- **Linear Integration**: All work is tracked as Linear issues, not local files Library RAG combine deux systèmes de recherche sémantique distincts:
- **Real-time Visibility**: Watch agent progress directly in your Linear workspace
- **Session Handoff**: Agents communicate via Linear comments, not text files
- **Two-Agent Pattern**: Initializer creates Linear project & issues, coding agents implement them
- **Initializer Bis**: Add new features to existing projects without re-initializing
- **Browser Testing**: Puppeteer MCP for UI verification
- **Claude Opus 4.5**: Uses Claude's most capable model by default
## Prerequisites 1. **📚 Library Philosophique** - Base documentaire de textes philosophiques (œuvres, chunks, résumés)
2. **🧠 Memory Ikario** - Système de mémoire conversationnelle (pensées et conversations)
### 1. Install Claude Code CLI and Python SDK **Architecture**: 5 collections Weaviate + GPU embedder (NVIDIA RTX 4070) + Mistral API
## 🏗️ Architecture
### Collections Weaviate (5)
```
📦 Library Philosophique (3 collections)
├─ Work → Métadonnées des œuvres philosophiques
├─ Chunk_v2 → 5355 passages de texte (1024-dim vectors)
└─ Summary_v2 → Résumés hiérarchiques des documents
🧠 Memory Ikario (2 collections)
├─ Thought → 104 pensées (réflexions, insights)
└─ Conversation → 12 conversations avec 380 messages
```
### GPU Embedder
- **Modèle**: BAAI/bge-m3 (1024 dimensions, 8192 tokens context)
- **GPU**: NVIDIA RTX 4070 Laptop (PyTorch CUDA + FP16)
- **Performance**: 30-70x plus rapide que Docker text2vec-transformers
- **Usage**: Vectorisation manuelle pour ingestion + requêtes
### Stack Technique
| Composant | Technologie | Rôle |
|-----------|-------------|------|
| **Vector DB** | Weaviate 1.34.4 | Stockage + recherche vectorielle |
| **Embeddings** | Python GPU embedder | Vectorisation (ingestion + requêtes) |
| **OCR** | Mistral OCR API | Extraction texte depuis PDF |
| **LLM** | Mistral Large / Ollama | Génération de réponses RAG |
| **Web** | Flask 3.0 + SSE | Interface web avec streaming |
| **Tests** | Puppeteer + pytest | Validation automatisée |
## 🚀 Démarrage Rapide
### 1. Prérequis
```bash ```bash
# Install Claude Code CLI (latest version required) # Python 3.10+
npm install -g @anthropic-ai/claude-code python --version
# Install Python dependencies # CUDA 12.4+ (pour GPU embedder)
nvidia-smi
# Docker (pour Weaviate)
docker --version
```
### 2. Installation
```bash
# Cloner le projet
git clone <repo-url>
cd linear_coding_library_rag
# Créer environnement virtuel
cd generations/library_rag
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
# Installer dépendances
pip install -r requirements.txt pip install -r requirements.txt
# PyTorch avec CUDA (si pas déjà installé)
pip install torch --index-url https://download.pytorch.org/whl/cu124
``` ```
### 2. Set Up Authentication ### 3. Configuration
Create a `.env` file in the root directory by copying the example:
```bash ```bash
# Copier le fichier d'exemple
cp .env.example .env cp .env.example .env
# Éditer .env avec vos clés API
nano .env
``` ```
Then configure your credentials in the `.env` file: **Variables requises**:
**1. Claude Code OAuth Token:**
```bash ```bash
# Generate the token using Claude Code CLI # Mistral API (OCR + LLM)
claude setup-token MISTRAL_API_KEY=your-mistral-api-key
# Add to .env file: # Ollama (optionnel, pour LLM local)
CLAUDE_CODE_OAUTH_TOKEN='your-oauth-token-here' OLLAMA_BASE_URL=http://localhost:11434
``` ```
**2. Linear API Key:** ### 4. Lancer les Services
```bash
# Get your API key from: https://linear.app/YOUR-TEAM/settings/api
# Add to .env file:
LINEAR_API_KEY='lin_api_xxxxxxxxxxxxx'
# Optional: Linear Team ID (if not set, agent will list teams)
LINEAR_TEAM_ID='your-team-id'
```
**Important:** The `.env` file is already in `.gitignore` - never commit it!
### 3. Verify Installation
```bash ```bash
claude --version # Should be latest version # Démarrer Weaviate
pip show claude-code-sdk # Check SDK is installed docker compose up -d
# Vérifier que Weaviate est prêt
curl http://localhost:8080/v1/.well-known/ready
# Lancer Flask
python flask_app.py
``` ```
## Quick Start **URLs**:
- 🌐 Flask: http://localhost:5000
- 🗄️ Weaviate: http://localhost:8080
### Option 1: Use the Example (Claude Clone) ## 📖 Utilisation
### Interface Web
Accéder à http://localhost:5000 pour:
| Page | URL | Description |
|------|-----|-------------|
| **Accueil** | `/` | Dashboard principal |
| **Recherche** | `/search` | Recherche dans library philosophique |
| **Chat** | `/chat` | Chat RAG avec contexte sémantique |
| **Memories** | `/memories` | Recherche dans pensées et messages |
| **Conversations** | `/conversations` | Historique des conversations |
| **Upload** | `/upload` | Ingestion de nouveaux PDF |
### 1. Recherche Philosophique
**Modes de recherche** (via `/search`):
- **📄 Simple**: Recherche directe dans les chunks
- **🌳 Hiérarchique**: Recherche par sections avec contexte
- **📚 Résumés**: Recherche dans les résumés de haut niveau
**Exemple**:
```
Requête: "la conscience selon Turing"
→ 16 résultats pertinents
→ Filtrage par auteur/œuvre
→ GPU embedder: ~17ms/requête
```
### 2. Chat RAG
**Fonctionnalités** (via `/chat`):
- 💬 Réponses longues et détaillées (500-800 mots)
- 📚 Citations directes des passages sources
- 🎯 Filtrage par œuvres (18 œuvres disponibles)
- 🔄 Streaming SSE (Server-Sent Events)
- 📖 Section "Sources utilisées" obligatoire
**Exemple de session**:
```
Question: "What is a Turing machine?"
→ Recherche sémantique: 11 chunks sur 5 sections
→ Génération LLM: ~30 secondes (Mistral Large)
→ Réponse académique détaillée avec sources
```
### 3. Memory Ikario
**Recherche dans pensées** (via `/memories`):
```
Requête: "test search"
→ 10 pensées pertinentes
→ Type: reflection, test, spontaneous
→ Concepts associés
```
**Recherche dans conversations**:
```
Requête: "philosophie intelligence"
→ Conversations pertinentes
→ Messages contextuels
→ Métadonnées (catégorie, date)
```
### 4. Ingestion de Documents
**Via interface web** (`/upload`):
1. Upload PDF (max 100 MB)
2. Sélection options:
- LLM provider (Mistral/Ollama)
- Chunking sémantique (optionnel)
- OCR annotations (optionnel)
3. Traitement automatique:
- OCR Mistral (~0.003€/page)
- Extraction métadonnées (auteur, titre, année)
- Chunking intelligent
- Vectorisation GPU (~15ms/chunk)
- Insertion Weaviate
**Via Python**:
```python
from utils.pdf_pipeline import process_pdf
result = process_pdf(
pdf_path="document.pdf",
use_llm=True,
llm_provider="mistral",
ingest_to_weaviate=True
)
print(f"Chunks: {result['chunks_count']}")
print(f"Cost: €{result['cost_total']:.4f}")
```
## 🧪 Tests
### Tests Automatisés
```bash ```bash
# Initialize the Claude Clone example project # Test ingestion GPU
python autonomous_agent_demo.py --project-dir ./ikario_body python test_gpu_mistral.py
# Add new features to an existing project # Test recherche sémantique (Puppeteer)
python autonomous_agent_demo.py --project-dir ./ikario_body --new-spec app_spec_theme_customization.txt node test_search_simple.js
# Test chat RAG (Puppeteer)
node test_chat_puppeteer.js
# Test memories/conversations (Puppeteer)
node test_memories_conversations.js
``` ```
For testing with limited iterations: **Résultats attendus**:
```bash - ✅ Ingestion: 9 chunks en ~1.2s
python autonomous_agent_demo.py --project-dir ./ikario_body --max-iterations 3 - ✅ Recherche: 16 résultats en ~2s
``` - ✅ Chat: 11 chunks, 5 sections, réponse complète
- ✅ Memories: API backend fonctionnelle
### Option 2: Create Your Own Application ### Tests Manuels
See the [Creating a New Application](#creating-a-new-application) section below for detailed instructions on creating a custom application from scratch.
## How It Works
### Linear-Centric Workflow
```
┌─────────────────────────────────────────────────────────────┐
│ LINEAR-INTEGRATED WORKFLOW │
├─────────────────────────────────────────────────────────────┤
│ app_spec.txt ──► Initializer Agent ──► Linear Issues (50) │
│ │ │
│ ┌─────────────────────────▼──────────┐ │
│ │ LINEAR WORKSPACE │ │
│ │ ┌────────────────────────────┐ │ │
│ │ │ Issue: Auth - Login flow │ │ │
│ │ │ Status: Todo → In Progress │ │ │
│ │ │ Comments: [session notes] │ │ │
│ │ └────────────────────────────┘ │ │
│ └────────────────────────────────────┘ │
│ │ │
│ Coding Agent queries Linear │
│ ├── Search for Todo issues │
│ ├── Update status to In Progress │
│ ├── Implement & test with Puppeteer │
│ ├── Add comment with implementation notes│
│ └── Update status to Done │
└─────────────────────────────────────────────────────────────┘
```
### Two-Agent Pattern
1. **Initializer Agent (Session 1):**
- Reads `app_spec.txt`
- Lists teams and creates a new Linear project
- Creates 50 Linear issues with detailed test steps
- Creates a META issue for session tracking
- Sets up project structure, `init.sh`, and git
2. **Coding Agent (Sessions 2+):**
- Queries Linear for highest-priority Todo issue
- Runs verification tests on previously completed features
- Claims issue (status → In Progress)
- Implements the feature
- Tests via Puppeteer browser automation
- Adds implementation comment to issue
- Marks complete (status → Done)
- Updates META issue with session summary
### Initializer Bis: Adding New Features
The **Initializer Bis** agent allows you to add new features to an existing project without re-initializing it. This is useful when you want to extend your application with additional functionality.
**How it works:**
1. Create a new specification file (e.g., `app_spec_theme_customization.txt`) in the `prompts/` directory
2. Run the agent with `--new-spec` flag pointing to your new spec file
3. The Initializer Bis agent will:
- Read the existing project state from `.linear_project.json`
- Read the new specification file
- Create new Linear issues for each `<feature>` tag in the spec
- Add these issues to the existing Linear project
- Update the META issue with information about the new features
- Copy the new spec file to the project directory
**Example:**
```bash
# Add theme customization features to an existing project
python autonomous_agent_demo.py --project-dir ./ikario_body --new-spec app_spec_theme_customization.txt
```
This will create multiple Linear issues (one per `<feature>` tag) that will be worked on by subsequent coding agent sessions.
### Session Handoff via Linear
Instead of local text files, agents communicate through:
- **Issue Comments**: Implementation details, blockers, context
- **META Issue**: Session summaries and handoff notes
- **Issue Status**: Todo / In Progress / Done workflow
## Configuration (.env file)
All configuration is done via a `.env` file in the root directory.
| Variable | Description | Required |
|----------|-------------|----------|
| `CLAUDE_CODE_OAUTH_TOKEN` | Claude Code OAuth token (from `claude setup-token`) | Yes |
| `LINEAR_API_KEY` | Linear API key for MCP access | Yes |
| `LINEAR_TEAM_ID` | Linear Team ID (if not set, agent will list teams and ask) | No |
## Command Line Options
| Option | Description | Default |
|--------|-------------|---------|
| `--project-dir` | Directory for the project | `./autonomous_demo_project` |
| `--max-iterations` | Max agent iterations | Unlimited |
| `--model` | Claude model to use | `claude-opus-4-5-20251101` |
| `--new-spec` | Name of new specification file to add (e.g., 'app_spec_new1.txt'). Use this to add new features to an existing project. | None |
## Project Structure
```
linear-agent-harness/
├── autonomous_agent_demo.py # Main entry point
├── agent.py # Agent session logic
├── client.py # Claude SDK + MCP client configuration
├── security.py # Bash command allowlist and validation
├── progress.py # Progress tracking utilities
├── prompts.py # Prompt loading utilities
├── linear_config.py # Linear configuration constants
├── prompts/
│ ├── app_spec.txt # Application specification (Claude Clone example)
│ ├── app_spec_template.txt # Template for creating new applications
│ ├── app_spec_theme_customization.txt # Example: Theme customization spec
│ ├── app_spec_mistral_extensible.txt # Example: Mistral provider spec
│ ├── initializer_prompt.md # First session prompt (creates Linear issues)
│ ├── initializer_bis_prompt.md # Prompt for adding new features
│ └── coding_prompt.md # Continuation session prompt (works issues)
└── requirements.txt # Python dependencies
```
## Generated Project Structure
After running, your project directory will contain:
```
ikario_body/
├── .linear_project.json # Linear project state (marker file)
├── app_spec.txt # Copied specification
├── app_spec_theme_customization.txt # New spec file (if using --new-spec)
├── init.sh # Environment setup script
├── .claude_settings.json # Security settings
└── [application files] # Generated application code
```
## MCP Servers Used
| Server | Transport | Purpose |
|--------|-----------|---------|
| **Linear** | HTTP (Streamable HTTP) | Project management - issues, status, comments |
| **Puppeteer** | stdio | Browser automation for UI testing |
## Security Model
This demo uses defense-in-depth security (see `security.py` and `client.py`):
1. **OS-level Sandbox:** Bash commands run in an isolated environment
2. **Filesystem Restrictions:** File operations restricted to project directory
3. **Bash Allowlist:** Only specific commands permitted (npm, node, git, etc.)
4. **MCP Permissions:** Tools explicitly allowed in security settings
## Linear Setup
Before running, ensure you have:
1. A Linear workspace with at least one team
2. An API key with read/write permissions (from Settings > API)
3. The agent will automatically detect your team and create a project
The initializer agent will create:
- A new Linear project named after your app
- 50 feature issues based on `app_spec.txt`
- 1 META issue for session tracking and handoff
All subsequent coding agents will work from this Linear project.
## Creating a New Application
This framework is designed to be **generic and reusable** for any web application. Here's how to create your own application from scratch.
### Understanding the Framework Structure
#### Generic Framework Files (DO NOT MODIFY)
These files work for all applications and should remain unchanged:
```
linear-coding-agent/
├── autonomous_agent_demo.py # Main entry point
├── agent.py # Agent session logic
├── client.py # Claude SDK + MCP client configuration
├── security.py # Bash command allowlist and validation
├── progress.py # Progress tracking utilities
├── prompts.py # Prompt loading utilities
├── linear_config.py # Linear configuration constants
├── requirements.txt # Python dependencies
└── prompts/
├── initializer_prompt.md # First session prompt template
├── initializer_bis_prompt.md # New features prompt template
└── coding_prompt.md # Continuation session prompt template
```
#### Application-Specific Files (CREATE THESE)
The **only file you need to create** is your application specification:
```
prompts/
└── app_spec.txt # Your application specification (XML format)
```
### Step-by-Step Guide
#### Step 1: Create Your Specification File
Create `prompts/app_spec.txt` using this XML structure:
```xml
<project_specification>
<project_name>Your Application Name</project_name>
<overview>
Complete description of your application. Explain what you want to build,
main objectives, and key features.
</overview>
<technology_stack>
<frontend>
<framework>React with Vite</framework>
<styling>Tailwind CSS</styling>
<state_management>React hooks</state_management>
</frontend>
<backend>
<runtime>Node.js with Express</runtime>
<database>SQLite</database>
</backend>
</technology_stack>
<prerequisites>
<environment_setup>
- List of prerequisites (dependencies, API keys, etc.)
</environment_setup>
</prerequisites>
<core_features>
<feature_1>
<title>Feature 1 Title</title>
<description>Detailed description</description>
<priority>1</priority>
<category>frontend</category>
<test_steps>
1. Test step 1
2. Test step 2
</test_steps>
</feature_1>
<feature_2>
<!-- More features -->
</feature_2>
</core_features>
</project_specification>
```
#### Step 2: Define Your Features
Each feature should have:
- **Title**: Clear, descriptive title
- **Description**: Complete explanation of what it does
- **Priority**: 1 (urgent) to 4 (optional)
- **Category**: `frontend`, `backend`, `database`, `auth`, `integration`, etc.
- **Test Steps**: Precise verification steps
Example feature:
```xml
<feature_1>
<title>User Authentication - Login Flow</title>
<description>
Implement authentication system with:
- Login form (email/password)
- Client and server-side validation
- JWT session management
- Password reset page
</description>
<priority>1</priority>
<category>auth</category>
<test_steps>
1. Access login page
2. Enter invalid email → see error
3. Enter valid credentials → redirect to dashboard
4. Verify JWT token is stored
5. Test logout functionality
</test_steps>
</feature_1>
```
#### Step 3: Launch Initialization
Once your `app_spec.txt` is ready:
```bash ```bash
python autonomous_agent_demo.py --project-dir ./my_new_app # Vérifier GPU embedder
curl http://localhost:5000/search?q=Turing
# Vérifier Weaviate
curl http://localhost:8080/v1/meta
# Vérifier nombre de chunks
python -c "import weaviate; c=weaviate.connect_to_local(); print(c.collections.get('Chunk_v2').aggregate.over_all()); c.close()"
``` ```
The initializer agent will: ## 📊 Métriques de Performance
1. Read your `app_spec.txt`
2. Create a Linear project
3. Create ~50 Linear issues based on your spec
4. Initialize project structure, `init.sh`, and git
#### Step 4: Monitor Development ### Ingestion
Coding agents will then: | Métrique | Avant (Docker) | Après (GPU) | Amélioration |
- Work on Linear issues one by one |----------|---------------|-------------|--------------|
- Implement features | **Vitesse** | 500-1000ms/chunk | 15ms/chunk | **30-70x** |
- Test with Puppeteer browser automation | **RAM** | 10 GB (container) | 0 GB | **-10 GB** |
- Update issues with implementation comments | **VRAM** | 0 GB | 2.6 GB | +2.6 GB |
- Mark issues as complete | **Architecture** | Hybride | Unifiée | Simplifiée |
### Minimal Example ### Recherche
Here's a minimal Todo App example to get started: | Opération | Temps | Détails |
|-----------|-------|---------|
| **Vectorisation requête** | ~17ms | GPU embedder (modèle chargé) |
| **Recherche Weaviate** | ~100-500ms | Selon complexité |
| **Recherche hiérarchique** | ~500ms | 11 chunks sur 5 sections |
| **Chat complet** | ~30s | Inclut génération LLM |
```xml ### Ressources
<project_specification>
<project_name>Todo App - Task Manager</project_name>
<overview> - **VRAM**: 2.6 GB peak (RTX 4070, 8 GB disponibles)
Simple web application for managing task lists. - **Modèle**: BAAI/bge-m3 (1024 dims, FP16 precision)
Users can create, edit, complete, and delete tasks. - **Batch size**: 48 (optimal pour RTX 4070)
</overview>
<technology_stack> ## 🔧 Configuration Avancée
<frontend>
<framework>React with Vite</framework>
<styling>Tailwind CSS</styling>
</frontend>
<backend>
<runtime>Node.js with Express</runtime>
<database>SQLite</database>
</backend>
</technology_stack>
<core_features> ### GPU Embedder
<feature_1>
<title>Main Interface - Task List</title>
<description>Display a list of all tasks with their status</description>
<priority>1</priority>
<category>frontend</category>
<test_steps>
1. Open application
2. Verify task list displays
</test_steps>
</feature_1>
<feature_2> **Fichier**: `memory/core/embedding_service.py`
<title>Create New Task</title>
<description>Form to add a new task to the list</description> ```python
<priority>1</priority> class GPUEmbeddingService:
<category>frontend</category> model_name = "BAAI/bge-m3"
<test_steps> embedding_dim = 1024
1. Click "New Task" optimal_batch_size = 48 # Ajuster selon GPU
2. Enter a title
3. Click "Add"
4. Verify task appears in list
</test_steps>
</feature_2>
</core_features>
</project_specification>
``` ```
### Best Practices **Réduire VRAM** (si Out of Memory):
```python
optimal_batch_size = 24 # Au lieu de 48
```
#### 1. Be Detailed but Structured ### Weaviate
Each feature must have: **Fichier**: `docker-compose.yml`
- Clear title
- Complete description of functionality
- Precise test steps
- Priority (1=urgent, 4=optional)
#### 2. Use Consistent XML Format ```yaml
services:
weaviate:
mem_limit: 8g # Limiter RAM
cpus: 4 # Limiter CPU
```
Follow the structure shown above for all features using `<feature_X>` tags. ### LLM Chat
#### 3. Organize by Categories **Fichier**: `flask_app.py` (ligne 1272)
Group features by category: ```python
- `auth`: Authentication # Personnaliser le prompt système
- `frontend`: User interface system_instruction = """
- `backend`: API and server logic Vous êtes un assistant expert en philosophie...
- `database`: Models and migrations """
- `integration`: External integrations ```
#### 4. Prioritize Features ## 📚 Documentation
- **Priority 1**: Critical features (auth, database) ### Structure du Projet
- **Priority 2**: Important features (core functionality)
- **Priority 3**: Secondary features (UX improvements)
- **Priority 4**: Nice-to-have (polish, optimizations)
### Using the Claude Clone as Reference ```
generations/library_rag/
├── flask_app.py # Application Flask principale
├── schema.py # Schémas Weaviate (5 collections)
├── docker-compose.yml # Weaviate (sans text2vec-transformers)
├── requirements.txt # Dépendances Python
├── .env.example # Configuration exemple
├── utils/
│ ├── pdf_pipeline.py # Pipeline ingestion PDF
│ ├── weaviate_ingest.py # Ingestion GPU vectorization
│ ├── llm_metadata.py # Extraction métadonnées LLM
│ └── ocr_processor.py # Mistral OCR
├── memory/
│ └── core/
│ └── embedding_service.py # GPU embedder
├── templates/ # Templates HTML
└── static/ # CSS, JS, images
The Claude Clone example in `prompts/app_spec.txt` is excellent reference material: docs/
├── migration-gpu/ # Documentation migration GPU embedder
│ ├── MIGRATION_GPU_EMBEDDER_SUCCESS.md
│ ├── TESTS_COMPLETS_GPU_EMBEDDER.md
│ └── ...
└── project_progress.md # Historique développement
#### ✅ Elements to Copy/Adapt: tests/
├── test_gpu_mistral.py # Test ingestion
├── test_search_simple.js # Test recherche
├── test_chat_puppeteer.js # Test chat
└── test_memories_conversations.js # Test memories
```
1. **XML Structure**: Overall structure with `<project_specification>`, `<overview>`, `<technology_stack>`, etc. ### Documentation Détaillée
2. **Feature Format**: How to structure `<feature_X>` tags with all required fields
3. **Technical Details**: How to describe technology stack, prerequisites, API endpoints, database schema, UI specs
#### ❌ Elements NOT to Copy: - **[Migration GPU Embedder](docs/migration-gpu/MIGRATION_GPU_EMBEDDER_SUCCESS.md)** - Rapport de migration détaillé
- **[Tests Complets](docs/migration-gpu/TESTS_COMPLETS_GPU_EMBEDDER.md)** - Résultats de tous les tests
- **[Project Progress](docs/project_progress.md)** - Historique du développement
- **[CHANGELOG](CHANGELOG.md)** - Historique des versions
1. **Specific Content**: Details about "Claude API", "artifacts", "conversations" are app-specific ## 🐛 Dépannage
2. **Business Features**: Adapt features to your application's needs
### Checklist for New Application ### Problème: "No module named 'memory'"
- [ ] Create `prompts/app_spec.txt` with your specification **Solution**:
- [ ] Define `<project_name>` for your application ```python
- [ ] Write complete `<overview>` # Vérifier sys.path dans weaviate_ingest.py
- [ ] Specify `<technology_stack>` (frontend + backend) sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent))
- [ ] List all `<prerequisites>` ```
- [ ] Define all `<core_features>` with `<feature_X>` tags
- [ ] Add `<test_steps>` for each feature
- [ ] Launch: `python autonomous_agent_demo.py --project-dir ./my_app`
- [ ] Verify in Linear that issues are created correctly
## Customization ### Problème: "CUDA not available"
### Adding New Features to Existing Projects **Solution**:
1. Create a new specification file in `prompts/` directory (e.g., `app_spec_new_feature.txt`)
2. Format it with `<feature>` tags following the same structure as `app_spec.txt`
3. Run with `--new-spec` flag:
```bash ```bash
python autonomous_agent_demo.py --project-dir ./ikario_body --new-spec app_spec_new_feature.txt # Réinstaller PyTorch avec CUDA
pip uninstall torch
pip install torch --index-url https://download.pytorch.org/whl/cu124
``` ```
4. The Initializer Bis agent will create new Linear issues for each feature in the spec file
### Adjusting Issue Count ### Problème: "Out of Memory (VRAM)"
Edit `prompts/initializer_prompt.md` and change "50 issues" to your desired count. **Solution**:
```python
# Réduire batch size dans embedding_service.py
optimal_batch_size = 24 # Au lieu de 48
```
### Modifying Allowed Commands ### Problème: Weaviate connection failed
Edit `security.py` to add or remove commands from `ALLOWED_COMMANDS`. **Solution**:
```bash
# Vérifier que Weaviate est lancé
docker compose ps
## Troubleshooting # Vérifier les logs
docker compose logs weaviate
**"CLAUDE_CODE_OAUTH_TOKEN not found in .env file"** # Redémarrer si nécessaire
1. Run `claude setup-token` to generate a token docker compose restart
2. Copy `.env.example` to `.env` ```
3. Add your token to the `.env` file
**"LINEAR_API_KEY not found in .env file"** ### Problème: Recherche ne renvoie rien
1. Get your API key from `https://linear.app/YOUR-TEAM/settings/api`
2. Add it to your `.env` file
**"Appears to hang on first run"** **Solution**:
Normal behavior. The initializer is creating a Linear project and 50 issues with detailed descriptions. Watch for `[Tool: mcp__linear__create_issue]` output. ```bash
# Vérifier nombre de chunks dans Weaviate
python -c "import weaviate; c=weaviate.connect_to_local(); print(f'Chunks: {c.collections.get(\"Chunk_v2\").aggregate.over_all().total_count}'); c.close()"
**"Command blocked by security hook"** # Réinjecter les données si nécessaire
The agent tried to run a disallowed command. Add it to `ALLOWED_COMMANDS` in `security.py` if needed. python schema.py --recreate-chunk
```
**"MCP server connection failed"** ## 🔐 Sécurité
Verify your `LINEAR_API_KEY` in the `.env` file is valid and has appropriate permissions. The Linear MCP server uses HTTP transport at `https://mcp.linear.app/mcp`.
## Viewing Progress - `.env` dans `.gitignore` (ne jamais commit les clés API)
- API Mistral: Facturation par usage (~€0.003/page OCR)
- Weaviate: Pas d'authentification (dev local uniquement)
- Flask: Mode debug (désactiver en production)
Open your Linear workspace to see: ## 📈 Roadmap
- The project created by the initializer agent
- All 50 issues organized under the project
- Real-time status changes (Todo → In Progress → Done)
- Implementation comments on each issue
- Session summaries on the META issue
- New issues added by Initializer Bis when using `--new-spec`
## License ### Court Terme
- [ ] Monitorer performance GPU en production
- [ ] Benchmarks formels sur gros documents (100+ pages)
- [ ] Tests unitaires pour `vectorize_chunks_batch()`
MIT License - see [LICENSE](LICENSE) for details. ### Moyen Terme
- [ ] API REST complète (OpenAPI/Swagger)
- [ ] Support multi-utilisateurs avec authentification
- [ ] Export résultats (PDF, Word, citations)
### Long Terme
- [ ] Fine-tuning BGE-M3 sur corpus philosophique
- [ ] Support langues supplémentaires (grec ancien, latin)
- [ ] Clustering automatique des concepts philosophiques
## 🤝 Contribution
1. Fork le projet
2. Créer une branche (`git checkout -b feature/amazing`)
3. Commit (`git commit -m 'Add amazing feature'`)
4. Push (`git push origin feature/amazing`)
5. Ouvrir une Pull Request
## 📄 Licence
MIT License - voir [LICENSE](LICENSE) pour détails.
## 🙏 Remerciements
- **Weaviate** - Vector database
- **BAAI** - BGE-M3 embedding model
- **Mistral AI** - OCR et LLM API
- **Anthropic** - Claude pour développement assisté
---
**Généré avec**: Claude Sonnet 4.5
**Dernière mise à jour**: Janvier 2026
**Version**: 2.0 (GPU Embedder Migration)