AI Novel Writer

A multi-agent system designed to write long-form coherent novels by maintaining a structured, multi-tier memory of the world, characters, and plot.

Many thanks to Gemini and GPT for their help!

Note

The project is currently in a very early stage.

👉 Project Architecture | 👉 Flowchart | 👉 User Guide | 👉 Documents

Key Features

• Long-Term Consistency: Uses a hybrid database (SQLite + FAISS) to track "hard" facts (rules, relationships) vs. "soft" details (descriptions, atmosphere).
• Multi-Agent Collaboration: Specialized AI roles (Architect, Writer, Critic, Scanner) work together to plan, write, review, and archive the story.
• Dual Model Support: Supports Google Gemini and OpenAI Compatible APIs.
• Dedicated Embedding Service: Decoupled architecture allowing separate endpoints for Generation (LLM) and Embeddings (Vector DB).
• Continuous Writing: Automated loop to generate multiple chapters in sequence.
• Two-Stage Plot Planning: Builds both a high-level Novel Plot Outline and a Detailed Plot Outline before chapter planning.
• Unified Discussion Template: World/plot/guide/chapter-text discussions now share one normalized template and title format.
• Centralized i18n & Prompt Management: All AI instructions and UI strings are strictly separated into the i18n/ directory, supporting zh-CN and en natively.
• Intent-Gated Retrieval Chain: Planning/writing now enforce Intent Classifier -> SQLite Prefilter -> FAISS Retrieval -> Cross-Tier Alignment.
• Atomic Memory Transactions: Synchronized SQLite and FAISS commits with index cloning and automated rollback on BLOCKING conflicts.
• Deep Interruption Recovery: Exhaustive integrity validation of all runtime artifacts before resuming.
• Language Guard: Automatic confidence-based language detection and rewrite loop.
• Conflict Severity Levels: Conflicts are classified as BLOCKING or NON_BLOCKING for safer automation.

Installation

1. Prerequisites:

   • Python 3.10+
   • (Optional) A supported local LLM model (GGUF format) for Llama.cpp or an OpenAI-compatible server.
   • (Optional) A supported local Embedding model for Llama.cpp or an OpenAI-compatible server.

2. Install Dependencies:

   pip install -r requirements.txt

3. Configuration:

   • Open src/config.py.
   • Set LANGUAGE = "Chinese" or "English".
   • Configure role model routing and workflow controls.

1. Initialize Workspace

python src/main.py --init

Fill novel/Novel_Overview.md and start:

python src/main.py --start

2. Manual Workflow

• Plan: python src/main.py --plan 1
• Write: python src/main.py --write 1
• Scan: python src/main.py --scan 1
• Conflicts: python src/main.py --conflicts-triage

3. Continuous Writing (Auto Mode)

# Generate 5 chapters starting from Chapter 1
python src/main.py --auto 1 5

GPU Acceleration & Model Configuration

The system now supports separating the Generation Model from the Embedding Model for optimal local performance.