version = 1

[llm]
model = "ollama:gemini-3-flash-preview"
temperature = 1.0
max_tokens = 4000
timeout = 120

[index]
quantization = true
chunk_size = 2000
chunk_overlap = 100
embeddings_batch_size = 16  # 16 files per batch - table.add() every 16 files for better persistence
embeddings_max_tokens_per_batch = 100000  # Keep existing token limit
flush_frequency = 2  # Flush every 2 batches = every 32 files for coordinated persistence
require_git = true  # Require git repository for indexing
contextual_model = "ollama:gemini-3-flash-preview"
contextual_descriptions = true
contextual_batch_size = 20
mcp_index = false

[search]
max_results = 10
similarity_threshold = 0.55
top_k = 20
output_format = "markdown"
max_files = 10
context_lines = 3
search_block_max_characters = 400  # Maximum characters to display per code/text/doc block
graph_expansion = false  # GraphRAG file-level expansion before reranking (matrix driver toggles this)

[search.reranker]
enabled = true
model = "voyage:rerank-2.5"
# Number of candidates to fetch from vector search before reranking
top_k_candidates = 50
# Number of final results to return after reranking
final_top_k = 10

[embedding]
code_model = "voyage:voyage-4-large"
text_model = "voyage:voyage-4"

[graphrag]
enabled = true
use_llm = true

[graphrag.llm]
description_model = "ollama:minimax-m2.7"
relationship_model = "ollama:minimax-m2.7"

# AI processing batch size - how many files to analyze per AI call
# Smaller values reduce API costs and token usage, larger values improve efficiency
ai_batch_size = 16

# Maximum tokens per batch request to avoid exceeding model limits
# Controls the total token count across all files in a single batch
max_batch_tokens = 16384

# Timeout for batch AI requests in seconds
# Longer timeouts accommodate larger batches but may delay processing
batch_timeout_seconds = 60

# Whether to fallback to individual AI calls if batch processing fails
# Recommended: true for reliability, false for strict batch-only processing
fallback_to_individual = true

# Maximum content sample size sent to AI for analysis (in tokens)
# Controls token usage and context window management using existing token estimation
max_sample_tokens = 1500

# Confidence threshold for filtering AI-discovered relationships (0.0-1.0)
# Higher values = more selective, only high-confidence relationships
# Recommended: 0.8 for production, 0.7 for development/exploration
confidence_threshold = 0.6

# Weight assigned to architectural relationships discovered by AI
# Higher values make AI-discovered relationships more prominent in graph
architectural_weight = 0.9

# System prompt for AI architectural relationship discovery
# Pure system instructions without data embedding
relationship_system_prompt = """You are an expert software architect specializing in code analysis. Analyze the provided code files and identify meaningful ARCHITECTURAL relationships that go beyond simple imports.

Focus on these relationship types:
- 'imports': Module/package imports and dependencies
- 'implements': Interface implementation, trait implementation
- 'extends': Class inheritance, module extension
- 'calls': Function/method calls between modules
- 'uses': Utility usage, service consumption
- 'configures': Configuration setup, dependency injection
- 'factory_creates': Factory pattern instantiation
- 'observer_pattern': Event listening, callback registration
- 'strategy_pattern': Algorithm selection, behavior delegation
- 'adapter_pattern': Interface adaptation, wrapper usage
- 'architectural_dependency': High-level system dependencies

Respond with a JSON array of relationships. Each relationship must include:
- source_path: relative path of source file
- target_path: relative path of target file
- relation_type: one of the types listed above
- description: specific explanation of HOW the relationship works
- confidence: 0.0-1.0 confidence score (use 0.8+ for clear relationships)

Only include relationships with clear architectural significance. Avoid trivial imports."""

# System prompt for AI file description generation
# Pure system instructions - file data sent separately
description_system_prompt = """You are a senior software engineer analyzing code architecture. Provide a concise 3-5 sentence description of the file's ROLE and PURPOSE in the system.

Focus on:
- What architectural layer this file belongs to (API, business logic, data access, utilities, etc.)
- Its primary responsibility and how it contributes to the system
- Key patterns or architectural decisions it implements

Avoid listing specific functions/classes. Instead, describe the file's architectural significance and how it fits into the larger system design."""

[search.hybrid]
# Enable hybrid search combining vector similarity with keyword matching
enabled = true
# Weight for vector similarity signal (0.0-1.0)
default_vector_weight = 0.7
# Weight for keyword matching signal (0.0-1.0)
default_keyword_weight = 0.3
# RRF dampening constant k (Cormack et al. 2009).
rrf_k = 60.0
# Per-query deterministic vector/keyword weight tilt (identifier vs natural-language).
auto_weight = false
# Weight for keyword matches in path/filename
keyword_path_weight = 2.0
# Weight for keyword matches in content
keyword_content_weight = 1.0
# Weight for keyword matches in symbols (code blocks only)
keyword_symbols_weight = 2.5
# Weight for keyword matches in title (document blocks only)
keyword_title_weight = 3.0

[search.reasoning]
enabled = false
model = "deepseek:deepseek-v4-flash"
max_candidates = 25
final_top_k = 10
context_level = "full"
reasoning_weight = 2.0