# Postgres (pgvector). For a host-run API using the Compose database on port 5433.
# Docker Compose overrides this with db:5432 for the containerized API.
# For native PostgreSQL, use its own credentials and port (usually localhost:5432).
DATABASE_URL=postgres://riwaq:riwaq@localhost:5433/riwaq

# --- LLM providers (an agent picks one via its `provider` field) ---

# Anthropic Claude (provider: "anthropic").
ANTHROPIC_API_KEY=

# Any OpenAI-compatible endpoint (provider: "openai").
# Works with OpenAI, OpenRouter, Groq, Together, local Ollama/vLLM/LM Studio, etc.
OPENAI_API_KEY=
OPENAI_BASE_URL=https://api.openai.com/v1
#   OpenRouter:  https://openrouter.ai/api/v1
#   Groq:        https://api.groq.com/openai/v1
#   Ollama:      http://localhost:11434/v1

# Deployment-wide LLM defaults. An org can override these (per-org config); an
# agent can override the org. Anything left unset falls back to these.
LLM_DEFAULT_PROVIDER=anthropic
LLM_DEFAULT_MODEL=
#   empty → uses the provider's built-in default (claude-haiku-4-5-20251001 / gpt-4o-mini)

# --- Embeddings (pluggable, with an offline fallback) ---
# Provider: voyage | openai | local. If unset: 'voyage' when EMBEDDINGS_API_KEY is
# set, otherwise 'local' (in-process, offline, no key, no extra service).
# EMBEDDINGS_PROVIDER=local
EMBEDDINGS_API_KEY=
# For the 'openai' provider — any /v1/embeddings server (OpenAI, Ollama, LM Studio…):
# EMBEDDINGS_BASE_URL=https://api.openai.com/v1
# Model override (empty → provider default: voyage-3 / text-embedding-3-small / all-MiniLM-L6-v2)
# EMBEDDINGS_MODEL=
# Vector dimension — MUST match the model. Locked into the DB at first migration.
#   local all-MiniLM-L6-v2 = 384 (default) · voyage-3 = 1024 · openai-3-small via dimensions = any
EMBEDDING_DIM=384

# --- Safety / limits ---
# development | test | production. In production, set LLM_ALLOWED_HOSTS below.
NODE_ENV=development
# SSRF egress control for tenant-supplied LLM/embedding base URLs. Private,
# loopback, link-local and cloud-metadata addresses are ALWAYS blocked. Optionally
# restrict to an explicit provider allowlist (comma-separated hostnames):
# LLM_ALLOWED_HOSTS=api.openai.com,api.anthropic.com,openrouter.ai
# Permit http:// (not just https) for tenant LLM URLs — local dev only.
ALLOW_INSECURE_LLM_URLS=0
# Base64-encoded 32-byte key for encrypting tenant LLM API keys at rest
# (AES-256-GCM). Optional in development and REQUIRED in production:
#   openssl rand -base64 32
SECRET_ENCRYPTION_KEY=
# Trusted organization backends sign short-lived end-user identity tokens with
# this base64-encoded 32+ byte HMAC secret. Required in production.
END_USER_SIGNING_SECRET=
# Provider SDK clients are bounded and expire so rotated credentials leave memory.
LLM_CLIENT_CACHE_MAX=100
LLM_CLIENT_CACHE_TTL_SECONDS=300
# Max request body size in bytes (default 20 MB; must stay above the 15 MB upload cap).
MAX_BODY_BYTES=20971520
# Maximum graceful HTTP drain before remaining connections are forced closed.
SHUTDOWN_TIMEOUT_MS=15000
TELEGRAM_API_TIMEOUT_MS=10000
# Outbound long polling is enabled by default. Disable only on replicas that
# should serve HTTP without owning Telegram polling work.
TELEGRAM_POLLING_ENABLED=1
# Start a fresh short-term channel conversation after this much inactivity, or
# after this many user turns. Durable per-user memory survives either rotation.
CHANNEL_SESSION_IDLE_MINUTES=30
CHANNEL_SESSION_MAX_TURNS=20

# --- Redis-backed production controls ---
# Required in production; enables durable jobs, shared rate limits, and caches.
REDIS_URL=
CACHE_TTL_SECONDS=30
# Required in production; gates organization provisioning, management, and /metrics.
# Compose also passes it to the dashboard as RIWAQ_ADMIN_TOKEN.
ADMIN_TOKEN=
RATE_LIMIT_WINDOW_SECONDS=60
RATE_LIMIT_PER_ORG=120
RATE_LIMIT_SIGNUP_PER_IP=5
MAX_CONCURRENT_REQUESTS_PER_ORG=20
# Persistent tenant governance. Set 0 for token/spend limits only to disable that
# specific ceiling. Cost values are deployment-calibrated micro-currency units.
ORG_MAX_TOTAL_TOKENS=10000000
ORG_MAX_ESTIMATED_COST_MICROS=100000000
ORG_MAX_DOCUMENTS=10000
ORG_MAX_STORED_CHARS=100000000
COST_PER_MILLION_INPUT_TOKENS_MICROS=0
COST_PER_MILLION_OUTPUT_TOKENS_MICROS=0

# --- Retrieval / RAG tuning ---
# Drop retrieved chunks below this cosine similarity (0..1); 0 = keep all. A small
# floor stops off-topic questions surfacing irrelevant chunks as "citations".
RETRIEVAL_MIN_SIMILARITY=0.2
# Cap total characters of retrieved context injected into a prompt.
RETRIEVAL_CHAR_BUDGET=12000
# HNSW candidate-list size (pgvector default 40). Higher = better recall under the
# per-tenant KB filter, at some latency cost.
RETRIEVAL_HNSW_EF_SEARCH=100
# A question at least this cosine-similar to a topic centroid joins that cluster.
TOPIC_MATCH_THRESHOLD=0.6
# Chunk sizing (characters). Defaults fit the shipped local embedder's ~256-token
# cap so no chunk is silently truncated at embed time. Raise for long-context
# embedders (Voyage/OpenAI) to get fewer, larger chunks.
CHUNK_MAX_CHARS=1000
CHUNK_OVERLAP_CHARS=150
# Cap prior-turn history characters injected into a prompt (context-window guard).
HISTORY_CHAR_BUDGET=8000
# Max long-term memory rows kept per (agent, end user).
MEMORY_MAX_PER_USER=500

# --- Built-in user-scoped actions ---
AGENT_TOOLS_ENABLED=1
# 1–8 model tool rounds, followed by a final answer; at most 12 tool calls per turn.
AGENT_MAX_TOOL_ROUNDS=4

# --- Reviewed learning ---
# An up-voted question this cosine-similar to an existing learned-answer candidate
# can join a candidate only when the normalized answer also matches.
LEARNED_DEDUP_SIMILARITY=0.9
# A question whose best retrieved chunk was below this similarity counts as a
# knowledge gap in the learning report.
LEARNING_GAP_SIMILARITY=0.25
# Votes never auto-publish. Operators must review exact sanitized text and confirm sharing.
# PUT /organizations/learning is deprecated and accepts only autoPromoteThreshold: 0.

# --- Reminders / scheduler ---
# Run the due-reminder poller in this process (multi-node safe). Set 0 on nodes
# that shouldn't fire reminders.
REMINDER_SCHEDULER_ENABLED=1
REMINDER_POLL_INTERVAL_MS=30000
REMINDER_BATCH=50
REMINDER_MAX_ATTEMPTS=5
REMINDER_WEBHOOK_TIMEOUT_MS=10000
# Renewed while delivering; an expired claim is recoverable after a crash.
REMINDER_CLAIM_LEASE_MS=300000
# Optional passive extraction adds one background model call per completed turn.
# Explicit chat tools are the default; passive extraction skips missing times.
REMINDER_AUTO_EXTRACT=0
# Guardrails for auto-extracted reminders.
REMINDER_MAX_HORIZON_DAYS=1825
REMINDER_MAX_PER_USER=100
# Without an org webhook, delivery requires this agent’s connected private Telegram
# chat for the canonical end user. Group-only sessions are never used.
# The per-org webhook the scheduler posts fired reminders to is set via
# PUT /organizations/webhook (SSRF-validated; a signing secret is returned once).

# --- Database pool ---
DB_POOL_MAX=10
# Server-side statement timeout (ms); 0 disables. Stops a runaway query from
# pinning a pooled connection (migrations lift this for their own transaction).
DB_STATEMENT_TIMEOUT_MS=30000

PORT=3000

# --- Optional Next.js management console (http://localhost:3001) ---
# Create an organization first, then paste the one-time API key here. When these
# are empty, the web console shows first-run setup instructions instead.
RIWAQ_API_KEY=
# Browser-facing API URL shown by the dashboard setup screen.
RIWAQ_PUBLIC_API_URL=http://localhost:3000
# Generate with: openssl rand -hex 32
RIWAQ_DASHBOARD_TOKEN=
WEB_PORT=3001