How idkhowtocodefr/vera-engine works
Vera Engine accepts category, merchant, customer, and trigger context, then processes available triggers into WhatsApp engagement actions. Its central workflow distills context, composes and validates a message through cloud-model fallbacks and a local model, and returns actions; a separate reply endpoint classifies incoming messages and chooses a response. The graph reflects the sampled API, composer, and model code plus README-documented behavior; unsampled implementation details are represented only where paths and capabilities are supported.
Main entry points: llm.py Model Adapters · composer.py Fallback Composer · app.py Trigger Processing · dataset Challenge Seed Data
Components 18
API and Context
- Context Ingestion
app.py - Context State
app.py - Trigger API
app.py - Reply API
app.py - Health and Metadata
app.py
Engagement Workflow
- Trigger Processing
app.py - Reply Classification
app.py
Message Composition
- Context Distiller
scripts/composer.py - Prompt Builder
scripts/prompts.py - Validation and Repair
scripts/composer.py
Model Inference
- Fallback Composer
scripts/composer.py - Model Adapters
scripts/llm.py - Local Qwen Model
scripts/llm.py
Other components
- Challenge Client
- NVIDIA NIM
- Groq
- Gemini
- Challenge Seed Data
dataset
Connections 23
- Challenge Client → Context Ingestion: pushes context
- Context Ingestion → Context State: stores payloads
- Challenge Client → Trigger API: submits triggers
- Trigger API → Context State: reads contexts
- Trigger API → Trigger Processing: dispatches
- Trigger Processing → Context State: loads context
- Trigger Processing → Context Distiller: composes message
- Context Distiller → Prompt Builder: builds local prompt
- Context Distiller → Fallback Composer: requests generation
- Fallback Composer → Model Adapters: tries model tiers
- Model Adapters → NVIDIA NIM: calls first
- Model Adapters → Groq: falls back
- Model Adapters → Gemini: falls back
- Model Adapters → Local Qwen Model: falls back
- Fallback Composer → Validation and Repair: validates and repairs
- Validation and Repair → Fallback Composer: returns result
- Trigger Processing → Trigger API: returns action
- Challenge Client → Reply API: posts reply
- Reply API → Reply Classification: classifies message
- Reply Classification → Fallback Composer: generates response
- Reply API → Challenge Client: returns decision
- Challenge Client → Health and Metadata: requests status
- Challenge Seed Data → Challenge Client: supplies test contexts