M.I.M.I.R - Multi-agent Intelligent Memory & Insight Repository

# NornicDB Configuration # # NornicDB is a high-performance drop-in replacement for Neo4j. # It receives pre-embedded nodes from Mimir - no embedding generation here. # # Environment variables take precedence over this file. # Use NORNIC_* prefix for environment variables. # ============================================================================= # Server Configuration # ============================================================================= server: # Bolt protocol port (Neo4j compatible) bolt_port: 7687 # HTTP API port (Neo4j compatible) http_port: 7474 # Data directory for persistence data_dir: ./data # Authentication (Neo4j compatible format: username:password) # Default: disabled (set to "none" or leave unset) # To enable: set to "admin:admin" or use environment: NEO4J_AUTH=admin/admin auth: "none" # disabled by default for easy development # Enable TLS (optional) tls: enabled: false cert_file: "" key_file: "" # ============================================================================= # GPU Configuration (NornicDB-specific) # ============================================================================= gpu: # Enable GPU acceleration for vector search enabled: true # Backend: auto, metal, cuda, opencl, vulkan, cpu # auto: Automatically detect best available backend # metal: Apple Silicon (M1/M2/M3) # cuda: NVIDIA GPUs # opencl: AMD GPUs (ROCm), some NVIDIA # vulkan: Cross-platform (experimental) # cpu: Fallback, no GPU backend: auto # Maximum VRAM usage (bytes, 0 = auto) # Default: ~80% of available VRAM max_vram_bytes: 0 # Batch size for GPU operations batch_size: 1000 # ============================================================================= # Memory Decay Configuration (NornicDB-specific) # ============================================================================= decay: # Enable memory decay system enabled: true # Run decay every N hours interval_hours: 24 # Memory types and their decay rates types: # Episodic: Personal events, experiences # Fast initial decay, stabilizes over time episodic: enabled: true initial_strength: 1.0 decay_rate: 0.1 # 10% per interval min_strength: 0.1 # Semantic: Facts, concepts, knowledge # Slow, steady decay semantic: enabled: true initial_strength: 1.0 decay_rate: 0.02 # 2% per interval min_strength: 0.3 # Procedural: Skills, how-to knowledge # Very slow decay procedural: enabled: true initial_strength: 1.0 decay_rate: 0.01 # 1% per interval min_strength: 0.5 # ============================================================================= # Auto-Relationship Configuration (NornicDB-specific) # ============================================================================= auto_relationships: # Enable automatic relationship creation enabled: true # Minimum similarity threshold for auto-linking (0.0-1.0) similarity_threshold: 0.85 # Maximum edges to auto-create per node max_edges_per_node: 10 # Types of auto-relationships to create types: # Link nodes with similar embeddings embedding_similarity: true # Link nodes accessed together co_access: true # Link nodes created close in time temporal_proximity: true # ============================================================================= # Topological Link Prediction (Feature Flag - NornicDB-specific) # ============================================================================= # Neo4j GDS-compatible topological link prediction # # IMPORTANT: Cypher procedures (CALL gds.linkPrediction.*) are ALWAYS available # This configuration only controls AUTOMATIC integration with inference engine # # Complements semantic inference with graph structure analysis topology_link_prediction: # Master feature flag - enables/disables AUTOMATIC topology integration # When false: Cypher procedures still work, but OnStore() won't use topology # When true: OnStore() automatically blends topology + semantic suggestions auto_integration_enabled: false # FEATURE FLAG: disabled by default (opt-in) # Algorithm to use for topology scoring # Options: adamic_adar, jaccard, common_neighbors, # resource_allocation, preferential_attachment, ensemble # adamic_adar: Best all-around for most graphs (recommended) # ensemble: Combines all 5 algorithms (slower, more robust) algorithm: adamic_adar # Top-K topology candidates to consider top_k: 10 # Minimum topology score threshold (0.0-1.0) # Lower = more suggestions, higher = fewer but more confident min_score: 0.3 # Weight for topology vs semantic (0.0-1.0) # 0.0 = pure semantic, 1.0 = pure topology # 0.4 = 40% topology, 60% semantic (recommended balanced) topology_weight: 0.4 # Graph cache refresh interval # Higher = better performance, lower = more up-to-date # 0 = rebuild on every prediction (safest, slowest) # 100 = rebuild every 100 predictions (recommended) graph_refresh_interval: 100 # A/B Testing support # Allows running experiments comparing topology vs semantic-only ab_testing: # Enable A/B test mode (splits traffic) enabled: false # Percentage of requests to use topology (0-100) # 50 = 50% get topology, 50% get semantic-only # Used for gradual rollout and A/B testing topology_percentage: 50 # Log predictions for analysis (debug only) log_predictions: false # ============================================================================= # Search Configuration # ============================================================================= search: # RRF (Reciprocal Rank Fusion) settings rrf: # RRF constant (higher = more equal weight to lower ranks) k: 60 # Default weight for vector search (0.0-1.0) vector_weight: 0.6 # Default weight for BM25 search (0.0-1.0) bm25_weight: 0.4 # Use adaptive weights based on query length adaptive: true # Properties indexed for full-text search # Must match Mimir's Neo4j node_search fulltext index fulltext_properties: - content - text - title - name - description - path - workerRole - requirements # ============================================================================= # Parallel Execution Configuration (NornicDB-specific) # ============================================================================= # Parallel execution significantly improves query performance on large datasets # by distributing work across multiple CPU cores. # # Performance: 3-4x speedup on large node sets (tested with 100K nodes) # Recommended for production workloads with >1000 nodes parallel: # Enable parallel query execution globally # Disabling reverts to sequential (single-threaded) execution enabled: true # Maximum number of worker goroutines # 0 = auto (uses all available CPU cores, runtime.NumCPU()) # Set explicitly for container resource limits or testing max_workers: 0 # Minimum batch size before parallelizing # Below this threshold, sequential execution is used (overhead not worth it) # Tune based on your typical query sizes: # - 500: Aggressive parallelization (more overhead, better for large queries) # - 1000: Balanced (recommended default) # - 2000: Conservative (less overhead, only parallelize large queries) min_batch_size: 1000 # ============================================================================= # Memory Management Configuration (NornicDB-specific) # ============================================================================= # Fine-grained control over Go's memory management to balance # performance vs memory usage for different deployment scenarios. # # Environment variables (override these settings): # NORNICDB_MEMORY_LIMIT, NORNICDB_GC_PERCENT, NORNICDB_POOL_ENABLED, # NORNICDB_QUERY_CACHE_SIZE, NORNICDB_QUERY_CACHE_TTL memory: # Soft memory limit (GOMEMLIMIT) # Examples: "2GB", "512MB", "0" (unlimited) # Recommended: Set to 80% of container/system memory # For 4GB container: "3GB" limit: "0" # Unlimited by default # GC aggressiveness (GOGC) # 100 = default (GC when heap doubles) # 50 = more frequent GC (lower memory, slightly more CPU) # 20 = aggressive GC (lowest memory, more CPU overhead) # 200 = less frequent GC (higher memory, better throughput) # -1 = disable GC (use with caution!) gc_percent: 100 # Object pooling for reduced allocations # Reuses query result structures, reducing GC pressure # Recommended: true for high-throughput workloads pool_enabled: true # Maximum objects per pool (memory leak prevention) pool_max_size: 1000 # Query plan caching # Avoids re-parsing identical Cypher queries query_cache: enabled: true size: 1000 # Maximum cached query plans ttl: "5m" # Cache entry TTL # ============================================================================= # Logging Configuration # ============================================================================= logging: # Log level: debug, info, warn, error level: info # Log format: json, text format: text # Log file (empty = stdout only) file: ""