config.toml

# MemoryOps Configuration
# Secrets (API keys, passwords) are NEVER stored here.
# Always use environment variables for secrets (see .env.example).

[server]
host = "0.0.0.0"
port = 8080
request_timeout_secs = 30
# Allow tools/skills endpoint URLs to resolve to loopback/private IP addresses.
# Useful for development or private VPC deployments. Off by default.
allow_private_ips = false

[database]
# Connection string read from DATABASE_URL env var.
# These settings control the connection pool.
max_connections = 20
min_connections = 2
connect_timeout_secs = 5

[redis]
# Connection string read from REDIS_URL env var.
pool_size = 10

# ── Embedding Provider ───────────────────────────────────────────────────────
[embedding]
# Options: "fastembed" (local, default) | "openai"
provider = "fastembed"
model = "BAAI/bge-small-en-v1.5"

# Uncomment to use OpenAI embeddings (requires OPENAI_API_KEY env var)
# [embedding.openai]
# model = "text-embedding-3-small"

# ── LLM Provider ─────────────────────────────────────────────────────────────
[llm]
# Options: "ollama" (local, default) | "openai" | "anthropic"
provider = "ollama"
model = "llama3"
# ⚠️  Docker users: "localhost" does NOT reach the host machine from inside a
#     container.  Use "host.docker.internal" instead (works on Docker Desktop
#     for Mac/Windows; on Linux add --add-host=host.docker.internal:host-gateway
#     to your docker run / compose service, or use the host network mode).
#
#     Local dev (bare metal / WSL):  base_url = "http://localhost:11434"
#     Docker Desktop (Mac / Windows): base_url = "http://host.docker.internal:11434"
base_url = "http://host.docker.internal:11434"
timeout_secs = 3

# Ollama Cloud / hosted Ollama (e.g. https://api.ollama.ai):
# Set base_url to your cloud endpoint and uncomment the block below.
# api_key_env names the environment variable that holds your Bearer token —
# the secret itself is never stored in this file.
#
# [llm.ollama]
# api_key_env = "OLLAMA_API_KEY"

# Uncomment to use OpenAI (requires OPENAI_API_KEY env var)
# [llm.openai]
# model = "gpt-4o-mini"

# Uncomment to use Anthropic (requires ANTHROPIC_API_KEY env var)
# [llm.anthropic]
# model = "claude-3-haiku-20240307"

# ── OpenAI-Compatible Providers ───────────────────────────────────────────────
# These providers share the /chat/completions request/response format.
# Set provider = "openrouter", "huggingface", or "openai_compatible".
# Omit base_url for openrouter/huggingface to use the defaults injected automatically.
#
# OpenRouter example:
# [llm]
# provider = "openrouter"
# model = "openai/gpt-4o-mini"
# timeout_secs = 30
#
# [llm.openai_compatible]
# api_key_env = "OPENROUTER_API_KEY"
# headers = { "HTTP-Referer" = "https://your-app.example", "X-OpenRouter-Title" = "MemoryOps" }
#
# Hugging Face Inference Router example:
# [llm]
# provider = "huggingface"
# model = "meta-llama/Llama-3.1-8B-Instruct"
# timeout_secs = 30
#
# [llm.openai_compatible]
# api_key_env = "HF_API_KEY"
#
# Generic OpenAI-compatible endpoint (e.g. vLLM, LM Studio, LocalAI):
# [llm]
# provider = "openai_compatible"
# model = "my-model"
# base_url = "http://localhost:8000/v1"
# timeout_secs = 60
#
# [llm.openai_compatible]
# api_key_env = "MY_PROVIDER_API_KEY"   # omit section if no auth is needed

# ── Processing Pipeline ──────────────────────────────────────────────────────
[processor]
# Fast path: max concurrency for synchronous processing
fast_path_concurrency = 4

# Slow path: number of Redis consumer workers
slow_path_workers = 2

# Max retry attempts before sending a job to the DLQ
max_retries = 3

# DLQ entry TTL in days
dlq_ttl_days = 7

# Reprocess a stuck 'processing' event if no heartbeat/update occurs for this many seconds
processing_stale_threshold_secs = 600

# UTC hour for daily maintenance pass (decay/pruning/hard-delete)
maintenance_window_hour_utc = 2

# UTC hour for weekly Sunday promotion pass
decay_window_hour_utc = 3

# ── Promotion Pipeline ───────────────────────────────────────────────────────
[promotion]
# How often to run the promotion scheduler
cadence_minutes = 15

# Episodic events older than this are eligible for clustering
clustering_window_hours = 24

# Minimum mean importance score for a cluster to be promoted
promotion_threshold = 0.65

# ── Decay & Pruning ──────────────────────────────────────────────────────────
[decay]
# Pruning scheduler runs daily at this UTC hour
schedule_hour_utc = 2

# Batch size for pruning to avoid table locks
batch_size = 1000

[decay.semantic]
decay_rate = 0.98
prune_threshold = 0.10

[decay.episodic]
decay_rate = 0.95
prune_threshold = 0.10

# ── Rate Limiting ────────────────────────────────────────────────────────────
[rate_limit]
# Default limits per workspace per minute (configurable per workspace via API)
retrieve_rpm = 60
ingest_rpm = 300
api_rpm = 120
# High-frequency dashboard polling routes get their own bucket so they cannot
# exhaust the workspace's general API quota.
dashboard_rpm = 600

# ── Retrieval Engine ─────────────────────────────────────────────────────────
[retrieval]
# Cosine similarity threshold above which a candidate is considered a duplicate
dedup_threshold = 0.92

# Default token budget if not specified in request
default_token_budget = 4096

# Tokenizer: cl100k_base (GPT-4 compatible)
tokenizer = "cl100k_base"

# Default scoring weights — must sum to 1.0
[retrieval.weights]
semantic_similarity = 0.35
importance = 0.25
recency = 0.20
source_authority = 0.10
memory_type = 0.10

# Default source authority scores
[retrieval.source_authority]
github = 0.9
slack = 0.6
jira = 0.7
linear = 0.7

# ── Observability ────────────────────────────────────────────────────────────
[telemetry]
# Log format: "json" (production) | "pretty" (development)
log_format = "pretty"

# OTEL exporter: "none" | "otlp" | "prometheus"
otel_exporter = "none"

# Slow span warning threshold in milliseconds
slow_span_threshold_ms = 500

# Retrieval trace retention in days
trace_retention_days = 30