Orchestrator Architecture
If "TOON" is the language your app speaks, the Orchestrator is the Director who speaks it. We replace linear pipelines with a "Loop of Reasoning" that mimics a human film director checking their work.
The Core Concept: The "Showrunner" Loop
In a standard LLM script, you fire a prompt and hope for the best. In an Orchestrator (ReAct) model, we build a Circular Dependency. The Orchestrator (Showrunner) refuses to generate a video frame until the "Context State" is verified.
The ReAct Pattern for Video Consistency
The Architecture: "The Supervisor Pattern"
We implement a Multi-Agent Orchestrator. You don't just have "One Bot"; you have a virtual film crew managed by a Supervisor.
Narrative Extractor
The Scriptwriter
Role: Reads the novel chunk.
Tool: Text_Compressor.
Output: Raw Scene Description.
Continuity Guard
The Script Supervisor
Role: Checks TOON database.
Tool: TOON_Retriever.
Output: "Correction: Night time, Arjun bleeding."
Prompt Engineer
The Cinematographer
Role: Merges Script + Consistency Check.
Tool: Video_Prompt_Generator.
Output: Final Stable Diffusion Prompt.
Logic Flow: The State Graph
The Orchestrator code binds them together. It runs a loop that blocks generation until consistency is met.
Query TOON State
Implementation Details
The Orchestrator State
Using TypeScript interfaces to define the "Showrunner's Clipboard".
interface ShowrunnerState {
current_chunk: string;
toon_history: string; // The compressed context
visual_plan: string | null;
video_url: string | null;
retry_count: number;
}Why This Solves "Hallucination"
In standard LLM apps, the model forgets that a character lost their sword three scenes ago.
- Without Orchestrator: The model guesses.
- With Orchestrator: The "Continuity Guard" agent forces the "Cinematographer" agent to include "No Sword" in the negative prompt or description before the request is ever sent to the Video AI.