Forseti461 Feature Architecture: Modular Prompts with Opik Versioning

Today we completed a major architectural milestone: modular prompt management for Forseti461. Each feature now has its own versioned prompt in Opik, enabling independent optimization and A/B testing.

astuce

From a single monolithic prompt to a clean separation of concerns — each Forseti feature can now evolve independently while sharing a common persona.

The Challenge

Our previous setup had prompts scattered across multiple files:

• app/agents/forseti/prompts.py — Python templates
• app/processors/workflows/workflow_autocontribution.py — Duplicate prompts
• Manual prompts in Opik (with a typo: foreseti461)

This made it hard to:

• Track which prompt version was deployed
• Run A/B tests on specific features
• Optimize one feature without affecting others
• Share prompts with N8N workflows (via Vaettir MCP)

The Solution: Feature-Based Prompt Architecture

We reorganized Forseti into 4 distinct features, each with its own prompt:

┌─────────────────────────────────────────────────────────────────────┐
│                    FORSETI 461 PROMPT ARCHITECTURE                   │
├─────────────────────────────────────────────────────────────────────┤
│                                                                      │
│  SHARED SYSTEM PROMPT (Persona)                                      │
│  ┌────────────────────────────────────────────────────────────────┐ │
│  │  forseti461-system-persona                                      │ │
│  │  "You are Forseti, a vigilant assistant..."                     │ │
│  └────────────────────────────────────────────────────────────────┘ │
│                              │                                       │
│              ┌───────────────┼───────────────┐                      │
│              ▼               ▼               ▼                      │
│  ┌──────────────────┐ ┌──────────────────┐ ┌──────────────────┐    │
│  │ CHARTER          │ │ CATEGORY         │ │ WORDING          │    │
│  │ VALIDATION       │ │ CLASSIFICATION   │ │ CORRECTION       │    │
│  ├──────────────────┤ ├──────────────────┤ ├──────────────────┤    │
│  │ forseti461-user- │ │ forseti461-user- │ │ forseti461-user- │    │
│  │ charter-validation│ │ category-        │ │ wording-         │    │
│  │                  │ │ classification   │ │ correction       │    │
│  └──────────────────┘ └──────────────────┘ └──────────────────┘    │
│                                                                      │
└─────────────────────────────────────────────────────────────────────┘

What We Built Today

1. Centralized Prompt Registry

All prompts now live in app/prompts/ with a unified registry:

from app.prompts import get_registry

registry = get_registry()

# Get formatted messages for LLM chat API
messages = registry.get_messages(
    "forseti.charter_validation",
    title="Ma proposition",
    body="Je propose d'améliorer..."
)

The registry supports:

• JSON chat format (Opik-compatible with Mustache {{input.var}} syntax)
• Python text format (legacy fallback with {var} syntax)
• Automatic fallback chain: Local JSON → Local Python → Opik remote

2. Consistent Opik Naming Convention

We standardized all prompt names following Option B (descriptive without dates):

Local Name	Opik Name	Commit
forseti.persona	forseti461-system-persona	f73d1f4a
forseti.charter_validation	forseti461-user-charter-validation	4b187053
forseti.category_classification	forseti461-user-category-classification	efd48155
forseti.wording_correction	forseti461-user-wording-correction	f2bbedd1
forseti.batch_validation	forseti461-user-batch-validation	0e097ebb

Naming pattern: forseti461-{role}-{feature}

3. Feature Documentation

Created comprehensive FORSETI_AGENT.md with a 7-step procedure to add new features:

1. Define prompt in app/prompts/local/forseti.py
2. Create result model in app/agents/forseti/models.py
3. Implement feature class in app/agents/forseti/features/
4. Register in features/__init__.py and agent.py
5. Add to forseti_charter.json for Opik sync
6. Write tests
7. Optimize with Opik (optional)

4. Batch Validation as Experiments

We decided not to make batch validation a feature class. Instead, it's handled as Opik experiments — allowing us to:

• Track metrics across validation runs
• Compare performance over time
• A/B test different prompt versions on the same dataset

File Structure

app/prompts/
├── __init__.py              # Exports: get_registry, CATEGORIES
├── registry.py              # PromptRegistry with chat format support
├── constants.py             # CATEGORIES, CATEGORY_DESCRIPTIONS
├── opik_sync.py             # CLI: python -m app.prompts.opik_sync
├── optimizer.py             # Prompt optimization with opik-optimizer
└── local/
    ├── forseti_charter.json # Opik-synced prompts (chat format)
    ├── forseti.py           # Python prompts (fallback)
    └── autocontrib.py       # Auto-contribution prompts

Syncing Prompts to Opik

One command syncs all Forseti prompts:

python -m app.prompts.opik_sync --prefix forseti.

# Output:
# ✅ forseti.persona (commit: f73d1f4a)
# ✅ forseti.charter_validation (commit: 4b187053)
# ✅ forseti.category_classification (commit: efd48155)
# ✅ forseti.wording_correction (commit: f2bbedd1)
# ✅ forseti.batch_validation (commit: 0e097ebb)

Why This Matters

Independent Optimization

Now we can optimize each feature separately:

from app.prompts import optimize_forseti_charter

# Optimize just category classification
result = optimize_forseti_charter(
    dataset_name="forseti-category-training",
    optimizer_type="meta_prompt",
)

Version Tracking

Every prompt change is tracked in Opik with a commit hash. We can:

• Roll back to previous versions
• Compare performance across versions
• Link experiments to specific prompt commits

Vaettir MCP Integration (Future)

With prompts centralized, we're ready for Phase 3:

• Share prompts between Python app and N8N workflows
• Access via MCP tools: get_prompt("forseti.charter_validation")
• Single source of truth across all integrations

What's Next

Phase 4: Update ForsetiAgent to use registry.get_messages() for all LLM calls, enabling:

• Automatic prompt versioning in traces
• Chat format for better model performance
• Seamless fallback if Opik is unavailable

---

Summary

Metric	Before	After
Prompt locations	3+ files	1 registry
Opik prompts synced	2 (with typo)	5 (correct naming)
Feature documentation	None	7-step procedure
Tests	17 passing	17 passing

Branch: experiment/heavy_dowel_6216

Key files:

• app/prompts/local/forseti_charter.json — Chat format prompts
• app/prompts/registry.py — Unified prompt access
• docs/docs/app/FORSETI_AGENT.md — Feature documentation
• docs/docs/app/core/prompts.md — Architecture overview

---

Modular prompts, independent optimization, clean architecture.