graphrag-core 0.2.0

# GraphRAG Advanced Features Configuration Example
# This file demonstrates all advanced features from Phases 2-3
# Copy and modify this file to enable state-of-the-art GraphRAG capabilities

# ============================================================================
# BASIC CONFIGURATION
# ============================================================================

output_dir = "output"
chunk_size = 400
chunk_overlap = 50
approach = "semantic"  # or "algorithmic" or "hybrid"

# ============================================================================
# PHASE 1 FEATURES (Foundation)
# ============================================================================

[entities]
min_confidence = 0.5
use_gleaning = true
max_gleaning_rounds = 2

# Phase 1.1: Triple Reflection - Validate extracted relationships
enable_triple_reflection = true
validation_min_confidence = 0.7  # Minimum confidence to keep a triple (0.0-1.0)

# Phase 1.2: Temporal Fields - Enable temporal reasoning
# Temporal fields are automatically extracted when present in text
# No additional configuration needed

# Phase 1.3: ATOM Atomic Fact Extraction (Simplified)
use_atomic_facts = false  # Set to true for fine-grained fact extraction
max_fact_tokens = 400     # Maximum tokens per atomic fact

# ============================================================================
# PHASE 2 FEATURES (Retrieval Enhancements)
# ============================================================================

[advanced_features]

# Phase 2.1: Symbolic Anchoring (CatRAG-style)
# Automatically applied for conceptual queries (e.g., "What is love?")
[advanced_features.symbolic_anchoring]
min_relevance = 0.3              # Minimum relevance score for anchors (0.0-1.0)
max_anchors = 5                  # Maximum anchors to extract per query
max_entities_per_anchor = 10     # Maximum entities grounded per anchor

# Phase 2.2: Dynamic Edge Weighting
# Query-aware relationship weight adjustment
[advanced_features.dynamic_weighting]
enable_semantic_boost = true     # Boost relationships semantically similar to query
enable_temporal_boost = true     # Boost recent/relevant temporal relationships
enable_concept_boost = true      # Boost relationships matching query concepts
enable_causal_boost = true       # Boost strong causal relationships

# Phase 2.3: Causal Chain Analysis
# Multi-step causal reasoning (e.g., "What caused X to lead to Y?")
[advanced_features.causal_analysis]
min_confidence = 0.3             # Minimum confidence for causal chains (0.0-1.0)
min_causal_strength = 0.5        # Minimum causal strength to consider (0.0-1.0)
max_chain_depth = 5              # Maximum chain depth to search
require_temporal_consistency = true  # Require chronological ordering in chains

# ============================================================================
# PHASE 3 FEATURES (Advanced Optimizations)
# ============================================================================

# Phase 3.1: Hierarchical Relationship Clustering
# Multi-level relationship organization using Leiden algorithm
[advanced_features.hierarchical_clustering]
num_levels = 3                   # Number of hierarchy levels (2-5)
resolutions = [0.8, 1.0, 1.5]    # Resolution parameters (higher = more clusters)
min_cluster_size = 3             # Minimum relationships per cluster
generate_summaries = true        # Generate LLM summaries for clusters (requires Ollama)

# Phase 3.2: Graph Weight Optimization (DW-GRPO)
# Heuristic optimization of relationship weights based on query performance
[advanced_features.weight_optimization]
learning_rate = 0.05             # Learning rate for weight adjustments (0.01-0.5)
max_iterations = 20              # Maximum optimization iterations
slope_window = 5                 # Window size for slope calculation
stagnation_threshold = 0.01      # Minimum slope to avoid stagnation
use_llm_eval = true              # Use LLM for quality evaluation (requires Ollama)

# Objective weights (must sum to ~1.0)
[advanced_features.weight_optimization.objective_weights]
relevance = 0.4                  # Weight for relevance objective
faithfulness = 0.4               # Weight for faithfulness objective
conciseness = 0.2                # Weight for conciseness objective

# ============================================================================
# EMBEDDINGS CONFIGURATION
# ============================================================================

[embeddings]
model = "nomic-embed-text"       # Ollama embedding model
dimension = 768
api_endpoint = "http://localhost:11434"
batch_size = 32

# ============================================================================
# GRAPH CONFIGURATION
# ============================================================================

[graph]
max_connections = 10
similarity_threshold = 0.7
extract_relationships = true
relationship_confidence_threshold = 0.6

[graph.traversal]
max_depth = 3
strategy = "bfs"                 # or "dfs"
max_results = 50

# ============================================================================
# RETRIEVAL CONFIGURATION
# ============================================================================

[retrieval]
top_k = 10
search_algorithm = "hybrid"      # "vector", "graph", or "hybrid"

# ============================================================================
# OLLAMA CONFIGURATION
# ============================================================================

[ollama]
host = "http://localhost:11434"
model = "llama3.2"               # Model for relationship extraction and summaries
embedding_model = "nomic-embed-text"
timeout_seconds = 300
max_retries = 3

# ============================================================================
# PARALLEL PROCESSING
# ============================================================================

[parallel]
enabled = true
num_threads = 4
min_batch_size = 10
chunk_batch_size = 4
parallel_embeddings = true
parallel_graph_ops = true
parallel_vector_ops = true

# ============================================================================
# USAGE NOTES
# ============================================================================

# 1. Triple Reflection: Improves quality by validating relationships against source text
#    - Best for: High-precision applications where accuracy matters
#    - Cost: +30-50% processing time
#    - Enable: entities.enable_triple_reflection = true

# 2. Atomic Fact Extraction: Fine-grained fact extraction (ATOM-style)
#    - Best for: Scientific texts, detailed analysis
#    - Cost: +50-100% processing time
#    - Enable: entities.use_atomic_facts = true

# 3. Symbolic Anchoring: Better conceptual/abstract query handling
#    - Best for: "What is X?" philosophical/conceptual queries
#    - Cost: Minimal (only affects retrieval)
#    - Auto-enabled for conceptual queries

# 4. Dynamic Edge Weighting: Query-aware relationship scoring
#    - Best for: Complex queries requiring context-aware ranking
#    - Cost: Minimal (only affects retrieval)
#    - Enable: advanced_features.dynamic_weighting.*

# 5. Causal Chain Analysis: Multi-step causal reasoning
#    - Best for: "Why did X cause Y?" causal queries
#    - Cost: Moderate (only for causal queries)
#    - Enable: advanced_features.causal_analysis.*

# 6. Hierarchical Clustering: Multi-level relationship organization
#    - Best for: Large graphs needing structure
#    - Cost: One-time build cost
#    - Enable: Call build_relationship_hierarchy() after graph construction

# 7. Weight Optimization: Improve retrieval quality over time
#    - Best for: Production systems with test queries
#    - Cost: One-time optimization phase
#    - Enable: advanced_features.weight_optimization.*

# ============================================================================
# RECOMMENDED CONFIGURATIONS
# ============================================================================

# HIGH PRECISION (research, accuracy-critical):
# - enable_triple_reflection = true
# - use_atomic_facts = true
# - validation_min_confidence = 0.8
# - min_confidence = 0.7

# BALANCED (general purpose):
# - enable_triple_reflection = true
# - use_atomic_facts = false
# - All dynamic weighting enabled
# - Causal analysis enabled

# HIGH THROUGHPUT (large-scale, performance-critical):
# - enable_triple_reflection = false
# - use_atomic_facts = false
# - generate_summaries = false
# - use_llm_eval = false
# - Rely on dynamic weighting only

# CONCEPTUAL/PHILOSOPHICAL QUERIES:
# - Symbolic anchoring enabled (default)
# - max_anchors = 10
# - enable_semantic_boost = true

# CAUSAL REASONING:
# - require_temporal_consistency = true
# - min_causal_strength = 0.7
# - max_chain_depth = 7