atlas.conf.example

#!/bin/bash
#===========================================
# ATLAS Configuration
#===========================================
# Copy this file to atlas.conf and customize.
# All values have sensible defaults.
#
# Usage:
#   cp atlas.conf.example atlas.conf
#   vim atlas.conf  # edit as needed
#   ./scripts/install.sh
#===========================================

#-------------------------------------------
# Network Configuration
#-------------------------------------------
# Node IP address (set to "auto" to detect, or specify manually)
ATLAS_NODE_IP="auto"

# Kubernetes namespace for all ATLAS services
ATLAS_NAMESPACE="atlas"

# External NodePorts (how you access services from outside the cluster)
ATLAS_API_PORTAL_NODEPORT=30000
ATLAS_LLM_PROXY_NODEPORT=30080
ATLAS_LENS_NODEPORT=31144
ATLAS_DASHBOARD_NODEPORT=30001
ATLAS_LLAMA_NODEPORT=32735
ATLAS_SANDBOX_NODEPORT=30820

# Internal service ports (usually don't change these)
ATLAS_REDIS_PORT=6379
ATLAS_LLAMA_PORT=8000
ATLAS_LENS_PORT=8001
ATLAS_API_PORTAL_PORT=3000
ATLAS_SANDBOX_PORT=8020
ATLAS_DASHBOARD_PORT=3001
ATLAS_LLM_PROXY_PORT=8000

#-------------------------------------------
# Storage Paths
#-------------------------------------------
# CHANGE THIS: Where models are stored (see docs/SETUP.md)
ATLAS_MODELS_DIR="/opt/atlas/models"

# CHANGE THIS: Where persistent data is stored
ATLAS_DATA_DIR="/opt/atlas/data"

# Training data directory
ATLAS_TRAINING_DIR="/opt/atlas/data/training"

# LoRA adapters directory
ATLAS_LORA_DIR="/opt/atlas/models/lora"

# Projects storage (for RAG)
ATLAS_PROJECTS_DIR="/opt/atlas/data/projects"

#-------------------------------------------
# Persistent Volume Sizes
#-------------------------------------------
ATLAS_PVC_REDIS_SIZE="5Gi"
ATLAS_PVC_PROJECTS_SIZE="20Gi"
ATLAS_PVC_API_PORTAL_SIZE="5Gi"

#-------------------------------------------
# Model Configuration
#-------------------------------------------
# Main model filename (must exist in ATLAS_MODELS_DIR)
ATLAS_MAIN_MODEL="Qwen3-14B-Q4_K_M.gguf"

# Draft model for speculative decoding (leave empty to disable)
ATLAS_DRAFT_MODEL="Qwen3-0.6B-Q8_0.gguf"

# Context window size (tokens)
# V3 uses --parallel 1 with 16384 context; set higher if using --parallel 2
ATLAS_CONTEXT_LENGTH=40960

# GPU layers to offload (99 = all layers)
ATLAS_GPU_LAYERS=99

# Parallel inference slots
# V3 requires --parallel 1 (VRAM too tight for 2 draft KV contexts with self-embeddings)
ATLAS_PARALLEL_SLOTS=1

# Flash attention (recommended for better performance)
ATLAS_FLASH_ATTENTION=true

#-------------------------------------------
# Resource Limits
#-------------------------------------------
# GPU memory to request/limit
ATLAS_LLAMA_GPU_MEMORY="14Gi"

# CPU limits per service
ATLAS_LLAMA_CPU_LIMIT="4"
ATLAS_LLAMA_CPU_REQUEST="2"
ATLAS_LLAMA_MEMORY_LIMIT="16Gi"
ATLAS_LLAMA_MEMORY_REQUEST="8Gi"

# Other services (lighter weight)
ATLAS_SERVICE_CPU_LIMIT="2"
ATLAS_SERVICE_CPU_REQUEST="0.5"
ATLAS_SERVICE_MEMORY_LIMIT="2Gi"
ATLAS_SERVICE_MEMORY_REQUEST="512Mi"

#-------------------------------------------
# Authentication & Security
#-------------------------------------------
# JWT secret (CHANGE THIS IN PRODUCTION)
# Leave as "auto" to generate random secret at install
ATLAS_JWT_SECRET="auto"

# JWT token expiry (hours)
ATLAS_JWT_EXPIRY_HOURS=24

# First admin user email (leave empty for first-registered-is-admin)
ATLAS_ADMIN_EMAIL=""

# Default rate limit for new API keys (requests per minute)
ATLAS_DEFAULT_RATE_LIMIT=1000

# API key hash algorithm
ATLAS_KEY_HASH_ALGORITHM="sha256"

#-------------------------------------------
# Feature Flags
#-------------------------------------------
# Enable speculative decoding (requires draft model)
ATLAS_ENABLE_SPECULATIVE=true

# Enable continuous learning (nightly training)
ATLAS_ENABLE_TRAINING=true

# Enable RAG (codebase context)
ATLAS_ENABLE_RAG=true

# Enable provenance tracking
ATLAS_ENABLE_PROVENANCE=true

# Enable real-time dashboard
ATLAS_ENABLE_DASHBOARD=true

#-------------------------------------------
# Timeouts (seconds)
#-------------------------------------------
ATLAS_LLM_TIMEOUT=120
ATLAS_SANDBOX_TIMEOUT=60
ATLAS_TASK_TIMEOUT=300
ATLAS_HEALTH_CHECK_TIMEOUT=10

#-------------------------------------------
# RAG Configuration
#-------------------------------------------
# Maximum context budget (tokens)
ATLAS_RAG_CONTEXT_BUDGET=8000

# Number of results to retrieve
ATLAS_RAG_TOP_K=20

# Maximum files per project
ATLAS_RAG_MAX_FILES=10000

#-------------------------------------------
# Training Configuration
#-------------------------------------------
# Minimum rating to include in training data
ATLAS_TRAINING_MIN_RATING=4

# Validation pass threshold (percentage)
ATLAS_TRAINING_VALIDATION_THRESHOLD=66

# LoRA rank
ATLAS_LORA_RANK=8

# LoRA alpha
ATLAS_LORA_ALPHA=16

# Training schedule (cron format)
ATLAS_TRAINING_SCHEDULE="0 2 * * *"

#-------------------------------------------
# Ralph Loop Configuration
#-------------------------------------------
# Maximum retry attempts
ATLAS_RALPH_MAX_RETRIES=5

# Base temperature
ATLAS_RALPH_BASE_TEMP=0.7

# Temperature increment per retry
ATLAS_RALPH_TEMP_INCREMENT=0.1

# Maximum temperature
ATLAS_RALPH_MAX_TEMP=1.2

#-------------------------------------------
# Logging
#-------------------------------------------
# Log level (DEBUG, INFO, WARNING, ERROR)
ATLAS_LOG_LEVEL="INFO"

# Enable request logging
ATLAS_LOG_REQUESTS=true

#-------------------------------------------
# URLs (for external access, optional)
#-------------------------------------------
# Set these if using reverse proxy/ingress
ATLAS_EXTERNAL_URL=""
ATLAS_API_EXTERNAL_URL=""

#-------------------------------------------
# V3 Pipeline — Phase Toggles
#-------------------------------------------
# Enable/disable entire V3 phases
ATLAS_V3_PHASE1_ENABLED=true
ATLAS_V3_PHASE2_ENABLED=true
ATLAS_V3_PHASE3_ENABLED=true

#-------------------------------------------
# V3 Phase 1: Constraint-Driven Generation
#-------------------------------------------
# Budget Forcing (1C) — Controls thinking token budget per difficulty
ATLAS_V3_BUDGET_FORCING_ENABLED=true
ATLAS_V3_BUDGET_FORCING_DEFAULT_TIER="standard"
ATLAS_V3_BUDGET_FORCING_MAX_WAIT_INJECTIONS=3

# PlanSearch (1A) — Constraint-guided plan generation
ATLAS_V3_PLAN_SEARCH_ENABLED=true
ATLAS_V3_PLAN_SEARCH_NUM_PLANS=3

# DivSampling (1B) — Prompt diversity perturbations
ATLAS_V3_DIV_SAMPLING_ENABLED=true

#-------------------------------------------
# V3 Phase 2: Adaptive Compute Allocation
#-------------------------------------------
# Blend-ASC (2A) — Adaptive K allocation based on Lens energy
ATLAS_V3_BLEND_ASC_ENABLED=true
ATLAS_V3_BLEND_ASC_DEFAULT_K=3

# ReASC (2B) — Early stopping on easy/confident tasks
ATLAS_V3_REASC_ENABLED=true
ATLAS_V3_REASC_CONFIDENCE_THRESHOLD=-0.5
ATLAS_V3_REASC_ENERGY_THRESHOLD=0.10

# S* Tiebreaking (2C) — Edge-case candidate comparison
ATLAS_V3_S_STAR_ENABLED=true
ATLAS_V3_S_STAR_ENERGY_DELTA=1.0

#-------------------------------------------
# V3 Phase 3: Verified Iterative Refinement
#-------------------------------------------
# Failure Analysis (3A) — Categorize why candidates fail
ATLAS_V3_FAILURE_ANALYSIS_ENABLED=true

# Constraint Refinement (3B) — Refine constraints from failures
ATLAS_V3_CONSTRAINT_REFINEMENT_ENABLED=true

# PR-CoT Repair (3C) — Multi-perspective quick repair
ATLAS_V3_PR_COT_ENABLED=true
ATLAS_V3_PR_COT_MAX_ROUNDS=2

# Derivation Chains (3D) — Verified sub-problem decomposition
ATLAS_V3_DERIVATION_CHAINS_ENABLED=true

# Refinement Loop (3E) — Full analyze→refine→generate→test cycle
ATLAS_V3_REFINEMENT_LOOP_ENABLED=true
ATLAS_V3_REFINEMENT_LOOP_MAX_ITERATIONS=5
ATLAS_V3_REFINEMENT_LOOP_TIME_BUDGET_SEC=300

# Metacognitive Model (3F) — Category-specific failure patterns
ATLAS_V3_METACOGNITIVE_ENABLED=true

# ACE Playbooks (3G) — Evolving persistent context
ATLAS_V3_ACE_ENABLED=true

# Self-Test Generation (Internal Verification)
ATLAS_V3_SELF_TEST_ENABLED=true
ATLAS_V3_SELF_TEST_NUM_CASES=5
ATLAS_V3_SELF_TEST_MAJORITY_THRESHOLD=0.6

#-------------------------------------------
# V3 Phase 4: Lens Evolution
#-------------------------------------------
# Replay Buffer (4A-CL) — Domain-stratified experience replay
ATLAS_V3_REPLAY_BUFFER_ENABLED=true
ATLAS_V3_REPLAY_BUFFER_MAX_SIZE=5000
ATLAS_V3_REPLAY_BUFFER_REPLAY_RATIO=0.30

# EWC (4A-EWC) — Elastic Weight Consolidation
ATLAS_V3_EWC_ENABLED=true
ATLAS_V3_EWC_LAMBDA=1000.0

# Lens Feedback Loop — Online recalibration during benchmarks
ATLAS_V3_LENS_FEEDBACK_ENABLED=false
ATLAS_V3_LENS_FEEDBACK_RETRAIN_INTERVAL=50
ATLAS_V3_LENS_FEEDBACK_DOMAIN="LCB"

#-------------------------------------------
# V3 Benchmark Runner — Ablation & Telemetry
#-------------------------------------------
# Candidate selection strategy: lens, random, logprob, oracle
ATLAS_V3_SELECTION_STRATEGY="lens"

# Enable Lens online feedback (Phase 4 evolution during runs)
ATLAS_V3_ENABLE_FEEDBACK=false

#-------------------------------------------
# Cache Manager — Proactive llama-server memory management
#-------------------------------------------
# Two-tier system: slot erase (fast, no downtime) + pod restart (slow, ~90s)
# Run: python3 scripts/llama-cache-manager.py --daemon
ATLAS_CACHE_MANAGER_ENABLED=true
ATLAS_CACHE_MANAGER_SOFT_THRESHOLD_MB=8192
ATLAS_CACHE_MANAGER_HARD_THRESHOLD_MB=10240
ATLAS_CACHE_MANAGER_CHECK_INTERVAL_SEC=30
ATLAS_CACHE_MANAGER_ERASE_COOLDOWN_SEC=60
ATLAS_CACHE_MANAGER_RESTART_COOLDOWN_SEC=300
ATLAS_CACHE_MANAGER_WARMUP_ENABLED=true

#-------------------------------------------
# Advanced (usually don't change)
#-------------------------------------------
# Container registry (for pre-built images)
ATLAS_REGISTRY="localhost"

# Image tag
ATLAS_IMAGE_TAG="latest"

# K3s config path
ATLAS_KUBECONFIG="/etc/rancher/k3s/k3s.yaml"