# YAMS Configuration Template
# This file contains all available configuration options with their default values.
# Copy this file to ~/.config/yams/config.toml and modify as needed.
#
# File Location Strategy:
# - All data files (databases, indexes, WAL, objects) are nested under data_dir by default
# - Paths can be absolute or relative; relative paths are resolved from data_dir
# - Use ~ for home directory expansion (e.g., ~/.yams/data)

[version]
config_version = 3
migrated_from = 0
migration_date = "2025-01-16"

[core]
# Data storage directory - all other data files nest under this by default.
# ~ is expanded to the home directory. The built-in default is $XDG_DATA_HOME/yams
# (~/.local/share/yams) on Linux and macOS and %LOCALAPPDATA%\yams on Windows.
# A value here takes precedence over YAMS_DATA_DIR.
data_dir = "~/.local/share/yams"
# Storage engine: "local" or "distributed" (future)
storage_engine = "local"
# Enable Write-Ahead Logging for crash recovery
enable_wal = true
# Enable telemetry (anonymous usage statistics)
enable_telemetry = false
# Log level: "trace", "debug", "info", "warn", "error"
log_level = "info"

[auth]
# Path to Ed25519 private key for signing
private_key_path = "~/.config/yams/keys/ed25519.pem"
# Path to Ed25519 public key for verification
public_key_path = "~/.config/yams/keys/ed25519.pub"
# API keys for authentication (optional)
api_keys = []

[performance]
# Number of worker threads for parallel operations
num_worker_threads = 4
# Size of I/O thread pool
io_thread_pool_size = 2
# Maximum concurrent operations
max_concurrent_operations = 100
# Enable memory-mapped file access
enable_memory_mapping = false
# Total cache size in MB
cache_size_mb = 512

[storage]
# Storage engine type
engine = "local"
# Base path for storage (defaults to core.data_dir)
base_path = "~/.yams/data"
# Enable transparent compression
enable_compression = true
# Enable content deduplication
enable_deduplication = true
# Enable content-defined chunking
enable_chunking = true
# CAS object storage directories (relative to base_path/data_dir)
objects_dir = "objects"
staging_dir = "staging"

[storage.s3]
# S3 Engine Settings (used if storage.engine = "s3")
url = ""
region = "us-east-1"
endpoint = ""
access_key = ""
secret_key = ""
session_token = ""
use_path_style = false
# Behavior when S3 is unreachable: "strict" (error) or "local" (fallback to local engine)
fallback_policy = "strict"
# Local data dir to use when falling back; empty = config-level data_dir
fallback_local_data_dir = ""

[storage.s3.r2]
# Explicit Cloudflare R2 temporary-credential bootstrap mode.
# Keep auth_mode="direct" for normal S3 key/secret usage.
auth_mode = "direct"
api_token = ""
account_id = ""
parent_access_key_id = ""
permission = "object-read-write"
ttl_seconds = 3600

[chunking]
# Rolling hash window size
window_size = 48
# Minimum chunk size in bytes
min_chunk_size = 1024
# Maximum chunk size in bytes
max_chunk_size = 16384
# Target average chunk size
average_chunk_size = 8192
# Hash algorithm for chunk identification
hash_algorithm = "sha256"

[compression]
# Enable compression for new content
enable = true
# Compression algorithm. Zstandard is the only compressor.
algorithm = "zstd"
# ZSTD compression level (1-22)
zstd_level = 9
# Minimum size to consider for compression
chunk_threshold = 1024
# Always compress files larger than this
always_compress_above = 10485760 # 10MB
# Never compress files smaller than this
never_compress_below = 1024 # 1KB
# Compress content older than N days
compress_after_days = 1
# Maximum parallel compression operations
max_concurrent_compressions = 4
# Use background compression
async_compression = true
# Decompression cache entries
decompression_cache_size = 100
# Decompression cache size in bytes
decompression_cache_size_bytes = 536870912 # 512MB

[wal]
# Enable Write-Ahead Logging
enable = true
# Directory for WAL files (relative to data_dir if not absolute)
wal_directory = "wal"
# Maximum size per log file
max_log_size = 104857600 # 100MB
# Sync to disk every N entries
sync_interval = 1000
# Maximum time between syncs (ms)
sync_timeout_ms = 100
# Compress rotated log files
compress_old_logs = true
# Maximum open log files
max_open_files = 10
# Batch commits for performance
enable_group_commit = true

[vector_database]
# Enable vector database for embeddings
enable = true
# SQLite database path for vectors (relative to data_dir if not absolute)
database_path = "vectors.db"
# Table name for embeddings
table_name = "document_embeddings"
# Embedding dimension (new installs default to simeon-default / 1024; existing DBs keep their stored dim)
embedding_dim = 1024
# Enable periodic checkpoints
enable_checkpoints = true
# Operations between checkpoints
checkpoint_frequency = 1000
# Maximum batch size for operations
max_batch_size = 1000
# Default similarity threshold for search
default_similarity_threshold = 0.35
# Use in-memory database (for testing)
use_in_memory = false
# Auto-create database if missing
auto_create = true
# Auto-repair on corruption
auto_repair = true
# Maximum number of vectors in index (for large corpuses, increase this)
# Default: 100000. For very large repositories, use 500000 or 1000000
max_elements = 100000

[embeddings]
# Enable embedding generation
enable = true
# Embedding backend: "daemon" or "simeon" (default, in-process, model-free)
# Env override: YAMS_EMBED_BACKEND
backend = "simeon"
# Automatically generate for new documents
auto_generate = false
# Automatically trigger embeddings after AddDocument
auto_on_add = false
# Preferred embedding model sentinel for the default simeon backend
preferred_model = "simeon-default"
# Root directory where optional ONNX plugin models are resolved
model_path = "~/.yams/models"
# Path to tokenizer
tokenizer_path = "models/tokenizer.json"
# Maximum sequence length
max_sequence_length = 512
# Embedding dimension (new installs default to simeon-default / 1024; existing DBs keep their stored dim)
embedding_dim = 1024
# Batch size for generation
batch_size = 32
# Normalize embedding vectors
normalize_embeddings = true
# Use GPU acceleration if available
enable_gpu = false
# Number of threads (-1 for auto)
num_threads = -1
# Delay between generation batches (ms)
generation_delay_ms = 1000
# Track content hash for staleness detection
track_content_hash = true
# Model version identifier
model_version = "1.0.0"
# Embedding schema version
embedding_schema_version = 1
# Keep the optional embedding model resident in memory (ONNX plugin reads this)
keep_model_hot = true
# Unload an idle embedding model after this many seconds (0 = never unload)
model_idle_timeout = 300
# Preload model on startup
preload_on_startup = false
# Maximum memory for model cache (MB)
max_model_memory_mb = 1024

[embeddings.simeon]
# Simeon encoder presets:
# - fast:    output_dim=384  projection="achlioptas_sparse"  lower memory / fastest encode
# - default: output_dim=1024 projection="fwht"               promoted first-class default
# - quality: output_dim=1024 projection="fwht"               pair with [search.simeon_query_router]
#            and [search.simeon_prf] enabled=true once those levers land
ngram_mode = "char_and_word"
ngram_min = 3
ngram_max = 5
sketch_dim = 4096
output_dim = 1024
projection = "fwht"
l2_normalize = true

[embeddings.simeon.bm25]
enabled = true
variant = "sab_smooth"
subword_gamma = 5.0
max_corpus_docs = 200000

[embeddings.simeon.bm25.router]
# Per-query dispatch between BM25 variants (Atire vs SabSmooth) via
# simeon::QueryRouter. When enabled, the lexical backend builds both
# indexes and picks per-query from pre-retrieval features.
# Preset details: third_party/simeon/docs/router_design.md.
enabled = true
preset = "passE_scq0_clar3"

[embeddings.selection]
# Selection strategy: "ranked" (score-based) or "intro_headings" (intro + headings only)
strategy = "ranked"
# Chunk selection mode before embedding: "full", "budgeted", or "adaptive"
mode = "budgeted"
# Maximum chunks selected per document when mode != "full"
max_chunks_per_doc = 8
# Maximum total characters selected per document when mode != "full"
max_chars_per_doc = 24000
# Additional score boost for heading-like chunks in adaptive mode
heading_boost = 1.25
# Additional score boost for intro/early chunks in adaptive mode
intro_boost = 0.75

# Daemon service and plugins
[daemon]
# Enable daemon-managed model provider and plugin auto-loading
enable = true
# Auto-start daemon if not running when needed
auto_start = true
# Unix domain socket path for IPC. Leave unset to use the per-user default:
# $XDG_RUNTIME_DIR/yams-daemon.sock, else /tmp/yams-daemon-<uid>.sock.
# socket_path = "/tmp/yams-daemon.sock"
# PID file for daemon process management (default: next to the socket's runtime dir)
# pid_file = "/tmp/yams-daemon.pid"
# Connection timeout in milliseconds
connect_timeout_ms = 1000
# Request timeout in milliseconds
request_timeout_ms = 5000
# Auto-load plugins from standard directories (and YAMS_PLUGIN_DIR if set)
auto_load_plugins = true
# Optional explicit plugin directory (comment out to use defaults and YAMS_PLUGIN_DIR)
# plugin_dir = "/usr/local/lib/yams/plugins"
# Strict plugin root mode: only scan trusted/configured roots, skip built-in defaults.
# Environment override: YAMS_PLUGIN_DIR_STRICT=1
plugin_dir_strict = false
# Enforce plugin naming policy: "relaxed" (accept yams_*_plugin.* or libyams_*_plugin.*)
# or "spec" (require libyams_*_plugin.*). Environment override: YAMS_PLUGIN_NAME_POLICY.
plugin_name_policy = "relaxed"
# Worker threads for daemon (0 = auto)
worker_threads = 0
# Max memory in GB the daemon will attempt to use (0 = unlimited)
max_memory_gb = 0
# Log level for daemon background service (trace|debug|info|warn|error)
log_level = "info"
# Do not rebuild vector schema automatically when embedding dimensions change.
# Run an explicit foreground vector/embedding repair after choosing the target dimension.
auto_rebuild_on_dim_mismatch = false
# Auto-repair missing/corrupted chunks and embeddings in the background
auto_repair = true
# Batch size for auto-repair passes
auto_repair_batch_size = 16
# Prefer centralized tuning (TuningManager); set to true only if you need legacy in-loop tuner
# use_legacy_tuner = false

[daemon.instrumentation]
# Runtime instrumentation profile.
#
# Profiles:
# - "normal" / "off" : disable automatic instrumentation behavior.
# - "memory" / "heap" / "malloc" : reduce concurrent startup maintenance so macOS
#   MallocStackLoggingNoCompact captures the target workload instead of background
#   repair, lexical, and vector-index rebuild noise.
profile = "normal"
# Override the profile defaults only when debugging a specific subsystem.
# suppress_auto_repair = true
# suppress_simeon_lexical_build = true
# suppress_vector_index_build = true

[daemon.graph_prune]
# Enable graph prune for snapshot-scoped version nodes
enabled = false
# Keep latest N snapshot versions per canonical node
keep_latest = 3
# How often to run prune
interval_minutes = 60
# Initial delay after startup
initial_delay_minutes = 10

[daemon.maintenance]
# Idle-only maintenance run by the repair service (no clients connected, no resource pressure).
# How often vectors.db is checked for a VACUUM that reclaims its free pages (0 disables).
vector_vacuum_interval_hours = 24
# Session files (under $XDG_STATE_HOME/yams/sessions) nobody has opened or changed for this
# many days are deleted, except the current session, sessions watching a directory that still
# exists, and sessions that still own documents (0 disables).
session_expiry_days = 30

# Host-load awareness. While the machine is busy (Linux PSI "some avg10", else load average
# per CPU; macOS load average and memory pressure), the daemon defers background work:
# topology rebuilds, repair scans, snapshot backfill, vectors.db VACUUM and semantic backfill.
# Integrity checks, crash recovery and WAL checkpoints are never deferred.
[tuning.resource]
# host_pressure = true               # false ignores host load
# host_cpu_pressure_pct = 40         # 1..100; raise on shared build hosts
# host_io_pressure_pct = 30
# host_memory_pressure_pct = 10
# host_load_per_cpu = 1.5            # fallback when PSI is unavailable
# background_max_deferral_s = 300    # longest deferral before work runs anyway; 0 never defers

# Centralized tuning knobs (advanced): defaults are safe; uncomment to override
[tuning]
# Default tuning preset applied by the daemon. Options: "efficient", "balanced", "aggressive".
profile = "balanced"
# Max concurrent in-flight list requests admitted to the daemon
list_inflight_limit = 8
# How long list requests wait for admission before erroring (ms)
list_admission_wait_ms = 200
# Max concurrent in-flight grep requests
grep_inflight_limit = 1
# How long grep requests wait for admission before erroring (ms)
grep_admission_wait_ms = 20000
# backpressure_read_pause_ms = 5
# worker_poll_ms = 150
# idle_cpu_pct = 10.0
# idle_mux_low_bytes = 4194304
# idle_shrink_hold_ms = 5000
# pool_cooldown_ms = 500
# pool_scale_step = 1
# pool_ipc_min = 1
# pool_ipc_max = 32
# pool_io_min = 1
# pool_io_max = 32
# io_conn_per_thread = 8          # grow IO pool only when activeConn > ioThreads * value
# post_ingest_threads = 0         # 0 = auto (recommend ~12.5% CPU budget, hard max 4)
# post_ingest_queue_max = 1000    # bounded queue; producers block when full (backpressure)
# use_internal_bus_for_repair = false      # when true, queue repair embed jobs to internal bus
# use_internal_bus_for_post_ingest = false # when true, queue post-ingest tasks to internal bus
# stream_mux_very_high_bytes = 268435456   # 256 MiB
# stream_mux_high_bytes = 134217728        # 128 MiB
# stream_mux_light1_bytes = 8388608        # 8 MiB
# stream_mux_light2_bytes = 33554432       # 32 MiB
# stream_mux_light3_bytes = 67108864       # 64 MiB

# Post-ingest pipeline concurrency caps. Omit a key to let the daemon auto-tune from the CPU
# budget (see TuneAdvisor::postIngestBudgetedConcurrency). Env vars (YAMS_POST_*_CONCURRENT)
# take precedence over these config keys. Valid ranges match the setter clamps:
#   total_concurrent: 1..256 ; embed_concurrent: 1..32
#   extraction_concurrent, kg_concurrent: 1..64 ; entity_concurrent, title_concurrent: 1..16
[tuning.post_ingest]
# total_concurrent = 8
# embed_concurrent = 4
# extraction_concurrent = 3
# kg_concurrent = 6
# entity_concurrent = 2
# title_concurrent = 2

# Notes:
# - When [embeddings].enable = false, the daemon automatically disables the model provider
#   and plugin auto-load to reduce startup overhead.
# - You can override plugin discovery with environment variables:
#   * YAMS_PLUGIN_DIR: an exclusive directory to scan for plugins
#   * YAMS_PLUGIN_DIR_STRICT: when truthy, skip built-in default plugin roots
# - Trust list persistence file is <data_dir>/plugins.trust
# - To control client/daemon timeouts at runtime, you can set:
#   * YAMS_REQUEST_TIMEOUT_MS (sets both header/body timeouts in client)
#   * YAMS_HEADER_TIMEOUT_MS / YAMS_BODY_TIMEOUT_MS (client-specific)

[embeddings.cache]
# Enable embedding cache for queries
enable_query_cache = true
# Cache size (number of queries)
cache_size = 1000
# Cache TTL in seconds
cache_ttl_seconds = 3600
# Use LRU eviction
lru_eviction = true


[vector_index]
# Index type: "FLAT", "HNSW", "IVF_FLAT", "IVF_PQ", "LSH", "ANNOY"
type = "HNSW"
# Vector dimension
dimension = 384
# Distance metric: "COSINE", "L2", "INNER_PRODUCT"
distance_metric = "COSINE"
# HNSW: connections per node
hnsw_m = 16
# HNSW: construction accuracy
hnsw_ef_construction = 200
# HNSW: search accuracy
hnsw_ef_search = 50
# Normalize vectors before indexing
normalize_vectors = true
# Maximum number of vectors
max_elements = 1000000
# Save/load index from disk
enable_persistence = true
# Index file path
index_path = "vector_index.bin"
# Batch size for operations
batch_size = 1000
# Use SIMD instructions
use_simd = true
# Number of threads (0 for auto)
num_threads = 0
# Use delta index for updates
enable_delta_index = true
# Merge delta when it reaches this size
delta_threshold = 1000

[search]
# Default result limit
default_limit = 10
# Maximum allowed limit
max_limit = 100
# Enable search result caching
enable_cache = true
# Cache size (number of queries)
cache_size = 1000
# Cache TTL in minutes
cache_ttl_minutes = 60
# Reranker backend: "simeon" (default), "onnx", "colbert", or "auto"
reranker_backend = "simeon"
# Preferred reranker model (optional)
reranker_model = ""

[search.path_tree]
# Enable path-tree traversal to serve hierarchical workloads (prefix/semantic/diff)
enable = false
# Mode when enabled: "fallback" keeps baseline scans as a safety net, "preferred" makes the tree the default
mode = "fallback"

[search.hybrid]
# Enable hybrid search
enable = true
# Enable knowledge graph scoring
enable_kg = true
# Weight for vector similarity
vector_weight = 0.50
# Weight for keyword relevance
keyword_weight = 0.30
# Weight for KG entity similarity
kg_entity_weight = 0.10
# Weight for structural prior
structural_weight = 0.05
# Weight for BART classification similarity (when enhanced search enabled)
classification_weight = 0.05
# Number of vector results to retrieve
vector_top_k = 50
# Number of keyword results to retrieve
keyword_top_k = 50
# Final number of results
final_top_k = 10
# Fusion strategy: "LINEAR_COMBINATION", "RECIPROCAL_RANK", "CASCADE"
fusion_strategy = "LINEAR_COMBINATION"
# RRF constant (for RECIPROCAL_RANK)
rrf_k = 60.0
# Enable post-fusion rerank-style score replacement
enable_reranking = true
# Number of top fused results eligible for rerank-style passes
rerank_top_k = 20
# Enable query expansion
enable_query_expansion = false
# Number of expansion terms
expansion_terms = 5
# Weight for expansion terms
expansion_weight = 0.3
# Maximum KG neighbors to consider
kg_max_neighbors = 32
# Maximum KG hops
kg_max_hops = 1
# Time budget for KG scoring (ms)
kg_budget_ms = 20
# Normalize scores to [0,1]
normalize_scores = true
# Generate search explanations
generate_explanations = true
# Include debug scores
include_debug_scores = false
# Run searches in parallel
parallel_search = true

[search.topology]
# Product-default topology context around hybrid retrieval. Use "disabled" only as a scientific
# control or explicit operator override.
mode = "hybrid_assist"
# Observe the routed projection while preserving the global ANN result set.
vector_policy = "shadow"
# Minimum selected-cluster prefix when adaptive routing is enabled
min_clusters = 1
# Max number of clusters considered when routing a weak query
max_clusters = 2
# Cap ranked lexical documents used as sparse routing evidence
max_seed_documents = 32
# Maximum dense representatives evaluated per cluster (0 = all stored representatives)
representative_limit = 0
# Grow the selected prefix while route scores remain this close to the best (0 = fixed max_clusters)
adaptive_probe_score_gap = 0.0
# Abstain from narrowing when the selected/excluded score boundary is smaller (0 = disabled)
narrow_min_boundary_margin = 0.20
# Max number of candidate docs pulled from routed clusters
max_docs = 64
# Score boost applied to medoid documents in routed clusters
medoid_boost = 0.05
# Maximum additive topology evidence applied during fusion
evidence_weight = 0.02
# Sparse/dense route score blend (0 = sparse only, 1 = dense only)
sparse_dense_alpha = 0.5
# Reject routes below this score
min_route_score = 0.0
# RRF k used by hybrid fusion
rrf_k = 12.0
# Cluster-route scoring rule for SparseGuidedClusterRouter.
#   "current"        — persistence + 0.5*cohesion + matchedSeed + memberCount/40 (default)
#   "size_weighted"  — current * 1/(1+log(size)); de-emphasizes giant clusters
#   "seed_coverage"  — rank by fraction of query seeds covered by each cluster
route_scoring = "current"

[topology]
# Cluster-formation engine used by the background topology rebuild.
#   "connected" — reciprocal-edge connected components (default, baseline)
#   "louvain"   — modularity-greedy communities on the reciprocal-edge graph
#   "kmeans"    — embedding-space coarse quantizer
engine = "connected"

[topology.tuner]
# Phase G: per-corpus adaptive topology tuner (UCB1 multi-armed bandit over
# supported engines). Each successful rebuild produces an
# intrinsic-quality reward (singletonRatio, giantClusterRatio, gini, intraEdge);
# the bandit learns which arm fits the corpus shape best. No relevance labels
# required. Off by default.
#
# Tuning recipes — uncomment one block to activate, leaving the others as
# documentation. Enabling the tuner overrides the static engine above whenever
# an arm is pulled.
#
# off (default): static [topology] knobs win; bandit dormant.
enabled = false
#
# default: 10-min cooldown, 100-doc delta, plan-G reward weights.
# Best fit for steady-state corpora that grow modestly between rebuilds.
# enabled = true
# cooldown_minutes = 10
# doc_count_delta = 100
#
# aggressive: 30-min cooldown, 500-doc delta, lower singleton penalty.
# Best fit for fast-growing corpora where rebuilds are expensive (50k+ docs).
# enabled = true
# cooldown_minutes = 30
# doc_count_delta = 500

[topology.tuner.reward]
# Intrinsic-reward weights. Reward = clamp(1 - α·singleton - β·giant -
# γ·|gini-0.4| + δ·intraEdge, 0, 1). Defaults reflect the Phase G plan.
# Lowering alpha/beta accepts noisier clusterings; raising delta favors
# tighter intra-cluster cohesion at the cost of global structure.
alpha_singleton = 0.4
beta_giant_cluster = 0.4
gamma_gini_deviation = 0.1
delta_intra_edge = 0.1

[knowledge_graph]
# Database path for KG (relative to data_dir if not absolute)
# Defaults to {data_dir}/yams.db
db_path = "yams.db"
# Enable full-text search for aliases
enable_alias_fts = true
# Enable WAL mode
enable_wal = true
# Prefer exact alias matches
prefer_exact_alias_first = true
# Default query limit
default_limit = 1000
# Node cache capacity
node_cache_capacity = 10000
# Alias cache capacity
alias_cache_capacity = 50000
# Embedding cache capacity
embedding_cache_capacity = 10000
# Neighbor cache capacity
neighbor_cache_capacity = 10000

# External knowledge graph endpoints for taxonomy/ontology retrieval
endpoints = ["http://localhost:7200/repositories/yams"]
# Default taxonomy domains for classification
default_domains = ["computer_science", "business", "research"]
# Cache taxonomy for N minutes
taxonomy_cache_minutes = 60
# Maximum related concepts to fetch per query
max_related_concepts = 20
# Maximum concept relationship depth to explore
max_concept_depth = 2
# HTTP timeout for KG requests (milliseconds)
request_timeout_ms = 3000
# Enable automatic domain detection from document content
enable_domain_detection = true
# SPARQL endpoint authentication (if required)
auth_username = ""
auth_password = ""
# Enable concept relationship caching
enable_relationship_cache = true
# Relationship cache TTL (minutes)
relationship_cache_ttl_minutes = 120

[file_detection]
# Use libmagic if available
use_lib_magic = true
# Use built-in patterns
use_builtin_patterns = true
# Use custom patterns
use_custom_patterns = true
# Path to custom patterns file
patterns_file = ""
# Maximum bytes to read for detection
max_bytes_to_read = 512
# Cache detection results
cache_results = true
# Cache size
cache_size = 1000

[mcp_server]
# Enable MCP server
enable = false
# Transport: "stdio", "websocket"
transport = "stdio"
# WebSocket host
host = "localhost"
# WebSocket port
port = 8080
# WebSocket path
path = "/mcp"
# Use SSL/TLS
use_ssl = false
# Connection timeout (ms)
connect_timeout_ms = 5000
# Receive timeout (ms)
receive_timeout_ms = 30000
# Maximum message size
max_message_size = 1048576 # 1MB
# Enable WebSocket ping/pong
enable_ping_pong = true
# Ping interval (seconds)
ping_interval_s = 30

# Optional: directory for file-backed prompt templates (Markdown).
# Files named PROMPT-*.md will be exposed via `prompts/list` and `prompts/get`.
# Examples: PROMPT-search-codebase.md -> name: "search_codebase"
# If not set, defaults to XDG_DATA_HOME/yams/prompts or ~/.local/share/yams/prompts,
# and as a last resort uses ./docs/prompts when running from the repo root.
# prompts_dir = "/home/user/.local/share/yams/prompts"

[daemon.models]
# Maximum number of models to keep loaded in memory
max_loaded_models = 4
# Number of models to keep always loaded (hot pool)
hot_pool_size = 2
# Models to preload on daemon startup
preload_models = ["bart-large-mnli"]
# Load models only when first requested
lazy_loading = false
# Unload models after being idle for N seconds
model_idle_timeout_s = 300
# Enable GPU acceleration if available
enable_gpu = false
# Number of threads for ONNX inference
num_threads = 4
# Model eviction policy: "lru" (least recently used), "lfu", "fifo"
eviction_policy = "lru"

[daemon.resource_pool]
# Minimum ONNX sessions per model
min_sessions = 1
# Maximum ONNX sessions per model
max_sessions = 3
# Session idle timeout in seconds
idle_timeout_s = 300
# Validate session health before use
validate_on_acquire = true

[classification]
# Enable BART-based document classification for intelligent tagging
enable = false
# Path to BART-MNLI model
model_path = "~/.yams/models/bart-large-mnli"
# Confidence threshold for tag assignment
confidence_threshold = 0.7
# Maximum tags per document
max_tags_per_document = 5
# Maximum sequence length for classification
max_sequence_length = 1024
# Enable batch processing for efficiency
enable_batch_processing = true
# Batch size for classification
batch_size = 16
# Model inference timeout (milliseconds)
inference_timeout_ms = 5000
# Automatically classify documents during indexing
auto_classify_on_index = true
# Enable real-time classification during search
enable_search_time_classification = false

[search.hotzones]
# Enable hotzone-based result boosting
enable = false
# Decay factor for hotzone relevance over time (0.0-1.0)
decay_factor = 0.95
# Minimum access frequency to maintain hotzone
min_frequency = 3
# Hotzone decay interval (hours)
decay_interval_hours = 24
# Maximum boost factor for hot concepts
max_boost_factor = 2.0
# Enable persistence of hotzone data across restarts
enable_persistence = true
# Hotzone data file path (relative to data_dir if not absolute)
data_file = "hotzones.db"
# Background decay processing interval (minutes)
decay_processing_interval_minutes = 60
# Enable cross-user hotzone sharing
enable_shared_hotzones = false

[search.enhanced]
# Enable enhanced search pipeline with classification and KG
enable = false
# Weight for BART classification scoring in final results
classification_weight = 0.15
# Weight for knowledge graph expansion in final results
kg_expansion_weight = 0.10
# Weight for hotzone boosting in final results
hotzone_weight = 0.05
# Enable cross-domain concept discovery
enable_cross_domain = true
# Maximum concepts for expansion per search result
max_expansion_concepts = 10
# Enable search result re-ranking based on concept similarity
enable_concept_reranking = true
# Timeout for enhanced search pipeline (milliseconds)
enhanced_search_timeout_ms = 2000

[migrations]
# Automatically migrate on version change
auto_migrate = true
# Create backup before migration
backup_before_migrate = true
# Migration log file path
migration_log_path = "migrations.log"

[indexing]
# PBI-040: Synchronous FTS5 indexing threshold
# When adding <= sync_threshold documents, perform FTS5 indexing synchronously
# instead of queueing to PostIngestQueue. This eliminates the 5-10s async delay
# for grep on small batches, making results immediately available.
# Set to 0 to always use async indexing (original behavior).
# Default: 10 files
sync_threshold = 10

[downloader]
# Default concurrency for parallel connections
default_concurrency = 4
# Default chunk size in bytes
default_chunk_size_bytes = 8388608
# Per-connection timeout in ms
default_timeout_ms = 60000
# Follow HTTP redirects
follow_redirects = true
# Enable resume using HTTP Range/If-Range
resume = true
# Global rate limit (bytes/sec, 0 = unlimited)
rate_limit_global_bps = 0
# Per-connection rate limit (bytes/sec, 0 = unlimited)
rate_limit_per_conn_bps = 0
# Retry/backoff policy
max_retry_attempts = 5
retry_backoff_ms = 500
retry_backoff_multiplier = 2.0
retry_max_backoff_ms = 15000
# Default checksum algorithm when provided
checksum_algo = "sha256"
# Temporary file extension for staging
temp_extension = ".part"
# Store directly into CAS by default; no user path writes unless export requested
store_only = true
# Maximum allowed file size in bytes (0 = unlimited)
max_file_bytes = 0

[downloader.tls]
# Verify TLS certificates (recommended)
verify = true
# Path to CA bundle (empty = system default)
ca_path = ""

[downloader.proxy]
# Proxy URL (e.g., http://user:pass@host:port) or empty
url = ""

[cli.streaming]
# Enable streaming for large payloads on CLI <-> daemon IPC
enable = true
# Switch to streaming when payload exceeds this many bytes
payload_threshold_bytes = 262144
# Switch to streaming when first-byte latency exceeds this many ms
ttfb_threshold_ms = 150
# Per-chunk size on the streaming transport
chunk_size_bytes = 65536
# Maximum number of in-flight chunks
max_inflight_chunks = 4

[cli.pool]
# CLI <-> daemon client pool sizing and timeouts
max_clients = 8
min_clients = 1
acquire_timeout_ms = 5000
# 0 = unbounded waiter queue
max_waiters = 0
fair_queue = true
health_check_interval_ms = 30000
health_check_timeout_ms = 500
reconnect_backoff_base_ms = 100
reconnect_backoff_max_ms = 5000
reconnect_backoff_jitter_pct = 0.2

[repair]
# Online repair of missing embeddings / corrupted chunks while serving traffic
enable_online_repair = false
max_repair_concurrency = 1
repair_backoff_ms = 250
max_retries = 3