v0.2.7 - Local Model Support & E2E Test Improvements

Highlights

• ✅ All 13 e2e tests now pass with local models (nemotron/Ollama, 30-90s processing time)
• 📚 Comprehensive local models setup guide for privacy-focused deployments
• 🔧 Fixed critical validation issues affecting all LLM providers
• 🚀 Installation improvements from v0.2.4-v0.2.6

What's Changed

E2E Test Improvements (v0.2.7)

Fixed: E2E tests fail with local models (nemotron/Ollama)

• FalkorDB's RediSearch parser treats certain words in group IDs as syntax operators
• Changed TEST_SESSION_GROUP from session-e2e_* → secondary{timestamp} (alphanumeric-only)
• Avoids syntax errors on words like "session", "temp", "grp2", etc.

Fixed: E2E test timeouts too short for local LLM processing

• Increased test timeouts from 15-60s → 120-180s
• Simplified episode content from complex multi-entity scenarios → 1-2 entities ("Alice likes pizza")
• Added polling instead of fixed waits for session-scoped episode test
• Result: All 13 e2e tests pass with nemotron at 30-90s per episode

Local Model Support (v0.2.4)

Added: Local models setup guide (docs/LOCAL_MODELS_SETUP.md)

• Comprehensive guide for running Graphiti MCP server with fully local models
• Supports Ollama, vLLM, and other local LLM providers
• Privacy-focused deployments with zero cloud dependencies
• References upstream PR: getzep/graphiti#1227

Fixed: Episode collision race conditions

• Episode names now use randomUUID() instead of Date.now()
• Eliminates collision risk in concurrent memory_store/auto-capture operations

Fixed: UUID resolution timeout too short

• Increased from 90s → 240s (80 attempts × 3s)
• Accommodates slower local model inference (60-120s vs OpenAI's 10-37s)

Fixed: Group ID validation issues

• Models were interpreting "(default: configured group)" literally
• Added sanitizeGroupId() to detect and reject invalid values
• Affects all models: Groq, OpenAI, Anthropic, etc.

Installation Fixes (v0.2.4-v0.2.6)

v0.2.6: Gateway shows helpful errors when config is missing

• Config parser now accepts undefined/null from installer
• Clear error message shows exact JSON snippet to add

v0.2.5: Removed JSON Schema required fields

• Installer no longer fails with empty config
• Added tests to prevent schema regression

v0.2.4: Additional installation improvements

• Empty config: {} now accepted during install
• Fixed grpc-js unhandled rejections crashing gateway
• Fixed Docker Compose FalkorDB connectivity

Testing

• ✅ All 13 e2e tests pass with OpenAI (gpt-4o-mini)
• ✅ All 13 e2e tests pass with nemotron (local Ollama)
• ✅ Installation works with empty config
• ✅ Group ID sanitization prevents SpiceDB errors

Breaking Changes

None. All changes are backward compatible.

Full Changelog

v0.2.7: v0.2.3...v0.2.7

Merged PRs:

• #34: Installation fixes (v0.2.4-v0.2.6)
• #35: Local model support + E2E improvements (v0.2.4, v0.2.7)