Nova: A Production-Ready Multimodal Video Agentic Workflow Platform

Nova is a LangGraph-based multimodal video search and creation platform that integrates Agentic Search and conversational video editing. Its core mission is to transform unstructured media content such as long videos and live stream replays into searchable, interpretable, editable, and exportable structured intelligent assets, enabling a paradigm shift in content production from manual browsing and clipping to AI Agent-driven workflows. The platform supports uploading videos to trigger processing like ASR, OCR, and Embedding to build searchable clips, generates/modifies clipping plans via natural language dialogue, and adopts innovative designs such as state persistence, composite routing, and minimal state patching.

The Nova AI-Cut Agent Platform aims to solve the inefficiency problem of traditional video processing and transform unstructured media content into structured intelligent assets. Designed for production environments, it supports uploading videos to trigger media processing workflows, builds searchable clips from ASR, OCR, Caption, etc., and finally generates or modifies video clipping plans via natural language dialogue.

Nova integrates two major technical directions: 1. LangGraph-based multimodal Agentic Search (responsible for intent routing, hybrid retrieval, etc.); 2. Dialogue-driven video editing model (minimal state patch incremental update to avoid full reconstruction). The top-level coordinator is the LangGraph Coordinator Graph, which includes five modules: intent routing layer (composite routing mechanism), perception and retrieval subgraph (rejects open LLM reflection), editing planning subgraph (minimal change principle), media workflow control, and export/rendering control.

Heavy media processing uses a dependency-aware DAG design (e.g., ASR depends on audio extraction, SegmentBuilder depends on ASR/OCR, etc.) to ensure optimal task sequence and parallelism. The state persistence layer includes core entities such as AgentState and GlobalEditingState, following the principle that domain.models only define DTOs. Practical examples of composite routing include scenarios like retrieval+editing, pure export, and complete workflow, demonstrating the process execution logic.

Core design principles: Agent orchestration priority, state-driven, reject open reflection, minimal change, clear service boundaries, dependency-aware DAG. The project value realizes the paradigm shift of video processing from manual to intelligent, batch to interactive, black-box to interpretable, one-time to asset-based, providing content creators and media companies with an AI-native workflow migration path.

Nova uses mature open-source components to build production-grade infrastructure: Celery/Redis (asynchronous tasks), MinIO (object storage), OpenSearch (full-text retrieval), Qdrant/Milvus (vector retrieval), ModelGateway (LLM API abstraction layer), ensuring system stability and scalability.