# everos default configuration.
#
# Lookup order (later overrides earlier):
#   1. This file (shipped defaults; lowest priority)
#   2. ~/.everos/config.toml — user-level overrides (optional;
#      path is overridable via EVEROS_CONFIG_FILE)
#   3. .env file in the working directory
#   4. Environment variables — EVEROS_<SECTION>__<KEY>
#         e.g. EVEROS_SQLITE__BUSY_TIMEOUT_MS=10000
#   5. Programmatic init args (highest priority)
#
# `null` (omitted in TOML) means "use the Pydantic default declared in code".

[memory]
# memory-root is the single directory holding all persisted memory.
# `~` is expanded; the path is resolved when MemoryRoot is constructed.
root = "~/.everos"
# Effective timezone for date buckets and timestamps. Drives
# component.utils.datetime; this is the SOLE source — OS `TZ` is not
# read. Override via `EVEROS_MEMORY__TIMEZONE` env var if needed.
timezone = "UTC"

[api]
# HTTP server bind for ``everos server start``. Default ``127.0.0.1``
# keeps the API on loopback only — EverOS ships no built-in auth (see
# SECURITY.md threat model). Only set ``host = "0.0.0.0"`` after you
# have placed your own gateway / auth layer in front of the server.
# Override via EVEROS_API__HOST and EVEROS_API__PORT.
host = "127.0.0.1"
port = 8000

[sqlite]
# PRAGMA journal_mode  — WAL is the recommended high-concurrency mode.
journal_mode = "WAL"
# PRAGMA synchronous  — NORMAL is safe under WAL and ~2x faster than FULL.
synchronous = "NORMAL"
# PRAGMA foreign_keys — must be explicitly enabled per connection.
foreign_keys = true
# PRAGMA temp_store   — MEMORY keeps query intermediates in RAM (no IO impact
# on durability — only affects sort/group/temp-table calculation buffers).
temp_store = "MEMORY"
# PRAGMA busy_timeout — milliseconds to wait on a locked DB before erroring.
busy_timeout_ms = 5000
# PRAGMA journal_size_limit — cap WAL/journal at ~64 MB.
journal_size_limit_bytes = 67108864
# PRAGMA cache_size  — KB of page cache (per connection).
cache_size_kb = 2048

[lancedb]
# Read consistency interval in seconds.
#   omitted / null -> no consistency check (fastest reads)
#   0              -> strict (every read checks updates)
#   >0             -> eventual (interval seconds between checks)
# Uncomment to override:
# read_consistency_seconds = 5.0

[llm]
# Provider-agnostic OpenAI-protocol client config. Override via env:
#   EVEROS_LLM__MODEL, EVEROS_LLM__API_KEY, EVEROS_LLM__BASE_URL
# Or via a ``.env`` file next to the project root (auto-loaded).
model = "gpt-4o-mini"
# api_key = ""
# base_url = ""

[multimodal]
# Independent LLM for multimodal parsing (everalgo-parser); must accept
# image / pdf / audio image_url parts. Override via env:
#   EVEROS_MULTIMODAL__MODEL, EVEROS_MULTIMODAL__API_KEY, EVEROS_MULTIMODAL__BASE_URL
model = "google/gemini-3-flash-preview"
max_concurrency = 4
# api_key = ""
# base_url = ""
# file:// content-item support (read locally by EverOS, not everalgo).
# file_uri_allow_dirs: empty = allow any readable file (local-first default);
# list base dirs to confine reads when the API is exposed.
# file_uri_allow_dirs = ["/srv/uploads"]
# file_uri_max_bytes = 52428800   # 50 MiB cap per file:// asset

[embedding]
# OpenAI-compatible embedding endpoint. Override via env:
#   EVEROS_EMBEDDING__MODEL, EVEROS_EMBEDDING__API_KEY, EVEROS_EMBEDDING__BASE_URL
# model / api_key / base_url have no shipped defaults — must be set
# (env or user toml) before the embedding capability is used.
# model     = "Qwen/Qwen3-Embedding-4B"
# api_key   = ""
# base_url  = "https://api.example.com/v1"
timeout_seconds = 30.0
max_retries     = 3
batch_size      = 10
max_concurrent  = 5

[rerank]
# Rerank provider. Override via env:
#   EVEROS_RERANK__PROVIDER, EVEROS_RERANK__MODEL, EVEROS_RERANK__API_KEY,
#   EVEROS_RERANK__BASE_URL
# `provider` picks the request-shape:
#   - "deepinfra" -> POST {base_url}/{model} (DeepInfra inference API)
#   - "vllm"      -> POST {base_url}/rerank (OpenAI-compat rerank endpoint)
provider = "deepinfra"
# model     = "Qwen/Qwen3-Reranker-4B"
# api_key   = ""
# base_url  = "https://api.deepinfra.com/v1/inference"
timeout_seconds = 30.0
max_retries     = 3
batch_size      = 10
max_concurrent  = 5

[boundary_detection]
# Passed through to ``everalgo.BoundaryDetector.adetect``.
hard_token_limit = 65536
hard_msg_limit = 500

[search]
# Vector retrieval strategy when SearchMethod.VECTOR is selected.
#   "maxsim_atomic" (default): ANN over atomic_fact.vector (pool=top_k*20),
#       max-pool the per-fact cosine by parent memcell, then reverse-resolve
#       to episode rows. MaxSim over atomic facts; +0.6pp over the legacy
#       episode-vector path on LoCoMo, at the cost of one extra LanceDB scan.
#   "episode": single-vector ANN over episode.vector (legacy path).
# Override via EVEROS_SEARCH__VECTOR_STRATEGY.
vector_strategy = "maxsim_atomic"

[memorize]
# Conversation mode. Selects the boundary detector and which pipelines run:
#   "chat"  -> BoundaryDetector       + user_memory only
#   "agent" -> AgentBoundaryDetector  + user_memory + agent_memory
# A single service process serves one mode at a time; switching mode
# requires a restart. Override via EVEROS_MEMORIZE__MODE.
mode = "agent"

# Maximum wall-clock for one memorize() invocation while holding the
# per-session lock. On timeout the outer asyncio.timeout cancels the call
# and the lock auto-releases so subsequent concurrent /add on the same
# session aren't deadlocked. Covers boundary LLM + memcell writes +
# synchronous portion of pipeline dispatch.
# Override via EVEROS_MEMORIZE__SESSION_LOCK_TIMEOUT_SECONDS.
session_lock_timeout_seconds = 360.0