Lunark Engine Architecture
Designs Optimal Conversations Every Moment
When a user's message arrives, it goes through a sophisticated 10-stage processing. Identifies intent and emotions, searches for relevant information in 7-type memory, and determines response strategies appropriate to current relationship state.
Sophisticated 10-Stage Processing
Lunark Engine goes through 10 stages of processing when a user's message arrives.
Identifies intent and emotions, searches for relevant information in 7-type memory.
Determines response strategies appropriate to current relationship state to design optimal conversations.
Processing Pipeline
10-Stage Processing Pipeline
Input Processing, Intent Classification, Detection & Analysis, Memory Search, Push-Pull Decision, Context Composition, Response Generation, Quality Verification, Memory Recording, Output Packaging
Language detection, preprocessing
needs_lore, needs_relation, basic, etc.
Detect emotion, profile, promise, episode triggers
Hot/Warm/Cold priority-based search
Recommend push/pull strategy for situation
Integrate prompt + memory + relationship
LLM call
Check prohibited expressions, push-pull compliance, relationship appropriateness
Separate sync/async storage
Response + emotion clip mapping
Model-Agnostic Structure
Lunark Engine is designed not to be tied to a specific language model.
Models are replaceable components, with the core value lying in the relationship-based conversation structure built on top.
This enables flexible response to technological changes.
Replaceable Models
LLMs are designed as replaceable components.
Relationship-Based Conversation Structure
The core value lies in the relationship-based conversation structure.
Flexible Response to Tech Changes
Model-agnostic structure enables flexible response to technological changes.