Abhigyan-Shekhar/Waggle-mcp

What's New

• Benchmark harness: end-to-end WaggleAdapter connecting the graph engine to ConvoMem / MemBench runners with automated exact-match scoring and latency logging.
• LongMemEval integration: CLI-driven ingestion and retrieval evaluation against the official LongMemEval split — 81.6% R@5 held-out is the headline generalization number.
• Logging utilities: structured log helpers (logging_utils) for consistent, level-aware output across all subsystems.
• Evidence tracking: evidence.py records source provenance on stored nodes so reasoning chains are fully traceable.
• Observability stack: Grafana dashboard, Prometheus config, and Docker Compose overlay in deploy/observability/.
• Kubernetes manifests: production-grade deployment.yaml, network policy, external-secret, and certificate templates under deploy/kubernetes/.
• Operational runbooks: incident response, secret management, API-key rotation, and onboarding guides in docs/runbooks/.
• README: honest benchmark presentation (held-out number leads), audience guide (individual dev vs. team), visible edges warning in Quick Start.

Install

pip install waggle-mcp==0.1.7

Honest benchmark note

81.6% R@5 is the held-out LongMemEval number — not used during development. The full-split ceiling of 97.4% is a retrieval bound on the saved benchmark setup. Both are real; the held-out one is the honest generalization number.

Deduplication recall sits at 77.3% (zero false-positive merges maintained). Improving recall is the primary 0.1.8 focus.