RagLeap Core v0.4.0 - I Shipped 13 Commits in a Day to Make AI Employees Production-Ready

I just shipped RagLeap Core v0.4.0 after 13 commits in a single day. 480+ commits total, 12 stars, and now it's a production-ready AI Business OS. What is RagLeap Core? Self-hosted, MIT licensed AI Business OS — 46 autonomous AI Employees, Graph RAG (Neo4j), multi-channel, BYOK across 17 LLM providers + Ollama, 8 libs (Python + Java). GitHub: https://github.com/antonyrag/ragleap-core 🔥 What shipped in v0.4.0 (Sep 12) 1. Silent Failures Are Now Loud - fix(employees) Before: Embedding failures in memory system were failing silently. Your AI employee would just forget things. Now: We surface them. If embedding fails, you know WHY. python # Before: silent pass # After: explicit error surfacing raise EmbeddingFailure(reason, context) 2. Production RUNBOOK - Because Sh*t Happens Added RUNBOOK.md for: db-downneo4j-crash-loopbackup-failurecert-expiredSelf-hosted means YOU own ops. So I documented how to recover when things break. No more guessing at 2 AM. 3. Graph RAG Goes Cross-Chunk - feat(ragleap-graph) Implemented cross-chunk relation extraction (#154). Old: Entities linked only within same chunk New: Relations extracted ACROSS chunks. If Doc says "API rate limit" in page 1 and "429 error" in page 5, Neo4j now links them. This is real Graph RAG, not just vector search + buzzword. 4. Java Port - 2 More Modules Done EmbeddingConfig -> JavaConnectionPool (db.py) -> JavaWhy Java? Regulated industries (HIPAA) need Java stack. We're at 9/8 libs now. Yes, more than planned. 5. ragleap-observability is LIVE on PyPI 📦javascriptpip install ragleap-observabilityWith live-verified status, postgres-exporter fix (runAsUser=65534), and proper publish CI job (#332). Why v0.4.0 Matters Most RAG demos die in production because: Silent failures No ops story No observability v0.4.0 fixes all three. Honest truncation notices for streamed answers, RUNBOOK, PyPI observability. Philosophy: Guardrailed autonomy for regulated industries. AI employees should be autonomous BUT you should be able to debug them. Stack Vector: pgvectorGraph: Neo4j (gracefully degrades if down)LLM: BYOK - 17 providers + OllamaLanguages: Python + JavaLicense: MIT - fully self-hostedWhat's Next? Product Hunt launch this week50 stars goal (12 -> 50)Multi-hop graph traversal wired into retrievalIf you're building self-hosted AI, give it a star and tell me what breaks. I ship fixes same day - literally 13 commits yesterday. Try it:bashgit clone https://github.com/antonyrag/ragleap-core docker-compose up