MiroFish Offline Skill Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection. MiroFish Offline is a fully local multi agent swarm intelligence engine. Feed it any document (press release, policy draft, financial report) and it generates hundreds of AI agents with unique personalities that simulate public reaction on social media — posts, arguments, opinion shifts — hour by hour. No cloud APIs required: Neo4j CE 5.15 handles graph memory, Ollama serves the LLMs. Architecture Overview Backend : Flask + Python 3.11 Frontend : Vue 3 + Node 18 Graph DB : Neo4j CE 5.15 (bolt protocol) LLM : Ollama (OpenAI compatible /v1 endpoint) Embeddings : nomic embed text (768 dimensional, via Ollama) Search : Hybrid — 0.7 × vector similarity + 0.3 × BM25 Installation Option A: Docker (Recommended) Open http://localhost:3000 . Option B: Manual Setup 1. Neo4j 2. Ollama 3. Backend 4. Frontend Configuration ( .env ) Core Python API GraphStorage Interface The abstraction layer between MiroFish and the graph database: Building a Knowledge Graph from a Document Creating and Running a Simulation Querying Simulation Results Chatting with a Simulated Agent Hybrid Search on the Knowledge Graph Implementing a Custom GraphStorage Backend Flask App Integration Pattern Accessing Storage in a Flask Route REST API Reference Method Endpoint Description POST /api/graph/build Upload document, build knowledge graph GET /api/graph/:id Get graph entities and relationships POST /api/simulation/create Create simulation environment POST /api/simulation/run Execute simulation GET /api/simulation/:id/results Get posts, sentiment, metrics GET /api/simulation/:id/agents List generated agents POST /api/report/generate Generate ReportAgent analysis POST /api/agent/:id/chat Chat with a specific agent GET /api/search Hybrid search the knowledge graph Example: Build graph from document Example: Run a simulation Hardware Selection Guide Use Case Model VRAM RAM Quick test / dev qwen2.5:7b 6 GB 16 GB Balanced quality qwen2.5:14b 10 GB 16 GB Production quality qwen2.5:32b 24 GB 32 GB CPU only (slow) qwen2.5:7b None 16 GB Switch model by editing .env : Then restart the backend — no other changes needed. Common Patterns PR Crisis Test Pipeline Use Any OpenAI Compatible Provider Troubleshooting Neo4j connection refused Ollama model not found Out of VRAM Embeddings dimension mismatch Docker Compose: Ollama container can't access GPU Slow simulation on CPU Use qwen2.5:7b for faster (lower quality) inference Reduce agent count to 50–100 for testing Reduce simulation hours to 6–12 CPU inference with 7b model: expect ~5–10 tokens/sec Frontend can't reach backend Project Structure

mirofish-offline-simulation