TwistedCollab Overview

TwistedCollab is a fully self-hosted browser application that serves as a local-first AI research assistant. It uses multi-agent workflows and local large language models (LLMs) to provide cloud-free support for科研 tasks. Key features include:

• Integration with the TwistedPair rhetorical transformation engine
• RAG retrieval, literature review, and debate analysis capabilities
• Zero cloud dependency to ensure data privacy and local control

Its core mission is to create a unified workspace for managing research data, documents, and creative ideas without relying on external cloud services.

Background & Core Idea

TwistedCollab was developed by satoruisaka to meet personal research needs. The developer identified that cloud-based AI tools were uneconomical and insecure for handling large volumes of sensitive research data. The core philosophy is:

• Build a local workspace to manage data, documents, and creativity
• Leverage local LLMs for complex language processing tasks
• Eliminate cloud dependency to protect privacy and reduce costs

The project focuses on supporting iterative thinking and creative构思, which are central to科研 work.

Architecture & Tech Stack

TwistedCollab uses a browser + FastAPI server architecture:

• Frontend: Pure HTML/JS app (communicates via SSE and REST)
• Backend: Python 3.10+ with core components:
  • ChatManager (session lifecycle, prompt assembly)
  • RetrievalManager (FAISS semantic search + SQLite FTS5 full-text search)
  • WebSearchClient (Brave API, DuckDuckGo)
  • OllamaClient (local LLM calls via Ollama REST API)
  • TwistedPairClient (rhetorical engine integration)
  • Skill System (YAML-defined multi-agent workflows)

Supported local models: ministral-3:14b, gemma4:26b, qwen3-coder:30b, deepseek-r1:8b. Runs on RTX 5090 for 14B-30B models.

Key Functions

1. RAG & Multi-source Retrieval: Supports 12 local data sources (literature PDFs, personal works, uploads, etc.) with dual index (FAISS semantic + FTS5 keyword search).
2. Multi-agent Workflows: Collab Tab enables YAML-defined skills (literature search, review, document comment) executed in sub-processes with CPU resource limits.
3. TwistedPair Engine: 3 adjustable knobs:
   • MODE (6 types: INVERT_ER, SO_WHAT_ER, etc.)
   • TONE (5 styles: NEUTRAL, TECHNICAL, etc.)
   • GAIN (10 levels: 1-3 factual, 7-10 creative)
4. Ecosystem Integration: Works with TwistedCore, TwistedDebate, TwistedDraw, and other Twisted services.

Daily Usage Workflow

The developer's typical day with TwistedCollab:

1. Morning: Read NewsAgent and TwistedNews emails for global updates.
2. Launch TwistedCollab:
   • Search Tab: Brainstorm with LLM + real-time web/doc retrieval
   • Notes Tab: Record new ideas or review old notes
   • Session Tab: Resume previous creative discussions
   • Collab Tab: Run agent workflows (literature search, review)
   • Tools Tab: Use Twisted services for deep analysis
3. For new workflows: Code in VS Code (Python + YAML) then restart the server.

Technical Innovations

1. Fully Local: Zero cloud dependency for model inference and data storage (privacy保障).
2. Dual Index Retrieval: Combines FAISS semantic search and SQLite FTS5 for comprehensive recall.
3. TwistedPair Paradigm: Treats LLM output as adjustable signal processing (mode, tone, gain).
4. Modular Skills: YAML-defined workflows with sub-process execution and resource limits.
5. Ecosystem Design: Integrated with multiple Twisted services to form a local AI toolchain.

Use Cases & Insights

适用场景:

• Academic researchers (literature management, data privacy)
• Knowledge workers (personal knowledge base)
• Creative writers (diverse text styles)
• Local AI enthusiasts (custom workflows)

Insights:

• Local LLMs (14B-30B) on RTX5090 can support complex科研 tasks.
• TwistedPair's signal processing analogy offers new LLM interaction design ideas.
• The modular architecture provides a reference for building personal AI workspaces.