Skip to content

ROADMAP.md

Latest commit

 

History

History
86 lines (64 loc) · 3.76 KB

ROADMAP.md

File metadata and controls

86 lines (64 loc) · 3.76 KB
Raw

MiroFish-Offline Roadmap

Current State (v0.2.0)

Fully local fork running on Neo4j CE + Ollama. All Zep Cloud dependencies removed. Core pipeline works: upload text → build knowledge graph → entity extraction → simulation → report generation.

Near Term

v0.3.0 — Stability & Python Compatibility

  • Fix camel-oasis / camel-ai compatibility with Python 3.12+ (currently requires <3.12)
  • Add Docker Compose GPU auto-detection (fallback to CPU-only Ollama)
  • Connection resilience: auto-reconnect to Neo4j on transient failures
  • Add /api/status endpoint showing Neo4j connection state, Ollama model availability, and disk usage
  • Structured logging with JSON output option

v0.4.0 — Search & Retrieval Improvements

  • Tune hybrid search weights (currently 0.7 vector / 0.3 BM25) — make configurable per graph
  • Add graph-aware reranking: boost results connected to the query entity
  • Support multiple embedding models (e.g., mxbai-embed-large, bge-m3 for multilingual)
  • Implement edge-weight decay for temporal relevance in simulations

Mid Term

v0.5.0 — Multi-Model Support

  • Model router: assign different Ollama models to different tasks (fast model for NER, large model for reports)
  • Support vLLM and llama.cpp as alternative backends alongside Ollama
  • Add model benchmarking tool: compare NER/RE quality across models on the same seed text
  • Quantization-aware config: auto-select context window based on available VRAM

v0.6.0 — Enhanced Simulation

  • Real-time simulation dashboard with WebSocket updates
  • Agent memory persistence across simulation rounds (currently in-memory)
  • Custom agent archetypes: define personality templates beyond OASIS defaults
  • Multi-language simulation support (agents can interact in different languages)
  • Export simulation transcripts as structured JSON for external analysis

v0.7.0 — Graph Intelligence

  • Community detection (Louvain/Leiden) to auto-identify entity clusters
  • Graph visualization improvements: force-directed layout, filtering by entity type
  • Temporal graph: track how entity relationships evolve across simulation rounds
  • Graph diff: compare two simulation runs side-by-side

Long Term

v1.0.0 — Production Ready

  • Authentication & multi-user support
  • Graph versioning: snapshot and restore graph states
  • Plugin system for custom NER extractors, search strategies, and report templates
  • Comprehensive test suite (unit + integration + E2E)
  • Performance benchmarks: document throughput (texts/min) and latency per hardware tier
  • Helm chart for Kubernetes deployment

Beyond v1.0

  • Federation: connect multiple MiroFish instances to share entity knowledge
  • Fine-tuned local models specifically trained for NER/RE on social simulation data
  • Voice-driven interaction: talk to simulation agents via local Whisper + TTS
  • Mobile companion app for monitoring running simulations

Hardware Tiers

Tier RAM GPU VRAM Recommended Model Expected Performance
Minimal 8 GB — (CPU only) qwen2.5:3b Slow, basic NER quality
Light 16 GB 6-8 GB qwen2.5:7b Usable for small graphs
Standard 32 GB 12-16 GB qwen2.5:14b Good for most use cases
Power 64 GB 24+ GB qwen2.5:32b Full quality, fast

Contributing

This project is AGPL-3.0 licensed. Contributions welcome — especially around:

  • Python 3.12+ compatibility for CAMEL-AI / OASIS
  • Additional embedding model support
  • Simulation quality improvements
  • Documentation and tutorials in English

See GitHub Issues for current tasks.