# DiscoveryGram configuration.
# Copy to .env and fill in. See docs/CONFIGURATION.md for the full reference.

# =============================================================================
# Telegram (required)
# =============================================================================
TELEGRAM_BOT_TOKEN=
# Comma-separated allow-list. At least one id is required — an empty list would
# expose the bot to everyone. Send /whoami to the bot to discover your id.
TELEGRAM_ALLOWED_USER_IDS=
# Optional: restrict group chats too. Empty means every chat an allowed user writes in.
TELEGRAM_ALLOWED_CHAT_IDS=
# polling (default, works anywhere) | webhook (production, behind a reverse proxy)
TELEGRAM_MODE=polling
# Public HTTPS URL of the reverse proxy in front of the bot. Required for webhook mode.
TELEGRAM_WEBHOOK_URL=
# Telegram signs every delivery with this, so requests that reach the port from
# anywhere else are rejected before being parsed. Strongly recommended.
TELEGRAM_WEBHOOK_SECRET=
TELEGRAM_WEBHOOK_LISTEN=0.0.0.0
TELEGRAM_WEBHOOK_PORT=8081
TELEGRAM_WEBHOOK_PATH=telegram
TELEGRAM_PARSE_MODE=MarkdownV2

# =============================================================================
# NoteDiscovery (required)
# =============================================================================
# Base URL *with port*. From inside Docker, a host instance is host.docker.internal.
NOTEDISCOVERY_URL=http://host.docker.internal:8000
# Optional — an instance may run unauthenticated. Sent as X-API-Key.
NOTEDISCOVERY_API_KEY=
NOTEDISCOVERY_TIMEOUT=30
NOTEDISCOVERY_MAX_RETRIES=3
NOTEDISCOVERY_VERIFY_TLS=true
# rest | mcp. REST is primary: NoteDiscovery's MCP server exposes a strict subset
# of the REST API (no media upload, export, sharing or folder move/rename/delete).
NOTEDISCOVERY_TRANSPORT=rest

# GET /api/search has no server-side default cap, so we always send one.
SEARCH_DEFAULT_LIMIT=50
# Shortest query NoteDiscovery will act on. It is a server constant that the API
# does not expose (2 in 0.31.3); shorter queries are refused locally instead.
SEARCH_MIN_QUERY_LENGTH=2
TREE_CACHE_TTL_S=300
# Where /quick captures land: one note per day, under this folder.
INBOX_PATH=Inbox
# With this off, a note aimed at a folder that does not exist is refused rather
# than silently inventing a folder tree.
AUTO_CREATE_PARENTS=true

# =============================================================================
# MCP subprocess (only when NOTEDISCOVERY_TRANSPORT=mcp)
# =============================================================================
MCP_ENABLED=false
# docker = spawn `docker run` (requires /var/run/docker.sock mounted — root-equivalent
#          host access, opt in deliberately)
# local  = spawn `python -m mcp_server` from a vendored module
MCP_LAUNCH_MODE=docker
MCP_DOCKER_IMAGE=ghcr.io/gamosoft/notediscovery:latest
MCP_STARTUP_TIMEOUT_S=30

# =============================================================================
# LLM router (phase 5)
# =============================================================================
# Failover works on (provider, model) pairs. Each pair is retried
# LLM_RETRIES_PER_MODEL times, then the next MODEL of the same provider is
# tried; only when a provider's models run out does the next provider take
# over. `make check-env` prints the exact ladder your configuration produces.
#
#   LLM_CHAIN_CHAT=groq,ollama with GROQ_MODELS=a,b and OLLAMA_MODELS=c gives:
#   groq/a (x3) -> groq/b (x3) -> ollama/c (x3) -> give up
LLM_CHAIN_CHAT=groq,cerebras,openrouter,ollama
LLM_CHAIN_VISION=gemini,openrouter,ollama

# Retries against the SAME (provider, model) pair before moving to the next model.
LLM_RETRIES_PER_MODEL=3
LLM_BACKOFF_BASE_S=1.0
LLM_REQUEST_TIMEOUT_S=60
LLM_CIRCUIT_FAILURE_THRESHOLD=5
LLM_CIRCUIT_RESET_S=120
# 0 disables the per-user daily cap. It counts *requests*: one capture that
# fails over across four rungs is one call against the user's budget.
LLM_DAILY_CALL_LIMIT_PER_USER=100
# Burst limit per user, per rolling minute; 0 disables it. The daily cap bounds
# spend, this bounds rate. Keep it loose: one photo capture is several calls
# (vision, tidy, title, tags, summary), and a limit that refuses halfway through
# leaves a half-written draft.
LLM_USER_RATE_PER_MINUTE=20
# Provider calls in flight across the whole process; 0 disables the bound.
# Back-pressure: past this, requests queue instead of piling onto a provider
# that is already the slow part.
LLM_MAX_CONCURRENT_REQUESTS=8

# --- Generation behaviour ---------------------------------------------------
# Most tags a generated note may carry. A model asked for "some tags" will
# happily return twenty, and the vault's tag index is shared across every note.
GENERATED_TAGS_MAX=5
# Whether a generated note records the provider and model that made it, as an
# HTML comment in the body — greppable, but invisible when the note is rendered.
PROVENANCE_ENABLED=true
# How many notes /ask may read as context. Each one is a vault read and tokens
# in the prompt, so this bounds both cost and latency.
ASK_CONTEXT_NOTES=5

# --- Provider credentials and model preferences -----------------------------
# <P>_MODELS and <P>_VISION_MODELS are ordered, comma-separated lists.
# A provider with no key (except ollama) or no model for the task is skipped,
# and the reason is logged at startup.
#
# Every <P>_BASE_URL below is optional and shown at its built-in default: set
# one only to point at a proxy, a self-hosted gateway or a pinned API version.
# See docs/LLM_PROVIDERS.md for per-provider setup and quirks.

# OpenAI-compatible (/chat/completions).
NVIDIA_API_KEY=
NVIDIA_BASE_URL=https://integrate.api.nvidia.com/v1
NVIDIA_MODELS=
NVIDIA_VISION_MODELS=

OPENROUTER_API_KEY=
# OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
OPENROUTER_MODELS=
OPENROUTER_VISION_MODELS=

GROQ_API_KEY=
# GROQ_BASE_URL=https://api.groq.com/openai/v1
GROQ_MODELS=
GROQ_VISION_MODELS=

# Gemini speaks its own dialect: key in x-goog-api-key, model in the URL path.
GEMINI_API_KEY=
# GEMINI_BASE_URL=https://generativelanguage.googleapis.com/v1beta
GEMINI_MODELS=
GEMINI_VISION_MODELS=

# Workers AI puts the account id in the URL, so the key alone is not enough:
# cloudflare in a chain without CLOUDFLARE_ACCOUNT_ID is skipped at startup.
CLOUDFLARE_API_KEY=
CLOUDFLARE_ACCOUNT_ID=
# CLOUDFLARE_BASE_URL=https://api.cloudflare.com/client/v4
CLOUDFLARE_MODELS=
CLOUDFLARE_VISION_MODELS=

# Cerebras serves text-only models. It is dropped from the vision ladder even
# if CEREBRAS_VISION_MODELS is set, so there is no such variable here.
CEREBRAS_API_KEY=
# CEREBRAS_BASE_URL=https://api.cerebras.ai/v1
CEREBRAS_MODELS=

MISTRAL_API_KEY=
# MISTRAL_BASE_URL=https://api.mistral.ai/v1
MISTRAL_MODELS=
MISTRAL_VISION_MODELS=

# Puter has no chat-completions endpoint; everything goes through one RPC
# call. Its API carries no version guarantee, so put it last in a chain.
PUTER_API_KEY=
# PUTER_BASE_URL=https://api.puter.com
PUTER_MODELS=
PUTER_VISION_MODELS=

# Ollama needs no key. The adapter uses its OpenAI-compatible surface at /v1;
# a bare host is accepted and the suffix added for you.
OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODELS=
OLLAMA_VISION_MODELS=

# =============================================================================
# Sessions, cache and limits
# =============================================================================
SESSION_BACKEND=memory
REDIS_URL=
SESSION_TTL_S=3600
RESULTS_PAGE_SIZE=5
# Default window for /recent, in days. Override per call with /recent 30.
RECENT_DEFAULT_DAYS=7
TREE_PAGE_SIZE=10
LONG_NOTE_MODE=paged
# Telegram's Bot API caps file downloads at 20 MB. A larger file is refused
# before it is downloaded, so it costs no transfer at all.
MAX_UPLOAD_MB=20
# search | quick. What a plain message (no command) does: run a search, or
# capture into today's inbox note. With `quick`, a message is never also
# searched — a thought you meant to keep must not become a query.
DEFAULT_TEXT_ACTION=search

# =============================================================================
# Observability
# =============================================================================
LOG_LEVEL=INFO
LOG_FORMAT=json
HEALTH_PORT=8080
METRICS_ENABLED=false