# =============================================================================
# Wiki-LLM configuration example
# -----------------------------------------------------------------------------
# Copy this file to ".env" and adjust the values. Every variable has a sane
# default, so the out-of-the-box setup (ChromaDB + local embeddings, no auth)
# works without editing anything except installing the optional `local` extra.
# =============================================================================

# -----------------------------------------------------------------------------
# HTTP server
# -----------------------------------------------------------------------------
APP_HOST=0.0.0.0
APP_PORT=8000

# -----------------------------------------------------------------------------
# Authentication (OPTIONAL, DISABLED BY DEFAULT)
# -----------------------------------------------------------------------------
# Leave API_KEY empty to keep all endpoints open. When set, every REST endpoint
# and the MCP HTTP/SSE endpoint require either:
#   Authorization: Bearer <API_KEY>      or      X-API-Key: <API_KEY>
# API_KEY=change-me-super-secret

# -----------------------------------------------------------------------------
# Logging
# -----------------------------------------------------------------------------
# Logs always go to stderr (never stdout, which is reserved for the MCP stdio
# JSON-RPC framing). Set LOG_FILE to also/instead write to a file.
LOG_LEVEL=INFO
# LOG_FILE=data/wikillm.log

# -----------------------------------------------------------------------------
# Storage
# -----------------------------------------------------------------------------
DATA_DIR=data
COLLECTION_NAME=wikillm

# Optional: ingest this file/directory once at stdio startup (before serving).
# INGEST_ON_START=/data
# INGEST_ON_START_RECURSIVE=true

# Vector backend: "chroma" (default, embedded) or "qdrant".
VECTOR_BACKEND=chroma

# --- ChromaDB ---
CHROMA_PATH=data/chroma

# --- Qdrant ---
# Use either a local embedded path OR a remote server URL (URL wins if set).
QDRANT_PATH=data/qdrant
# QDRANT_URL=http://localhost:6333
# QDRANT_API_KEY=

# -----------------------------------------------------------------------------
# Chunking
# -----------------------------------------------------------------------------
CHUNK_SIZE=1000
CHUNK_OVERLAP=150

# -----------------------------------------------------------------------------
# Embeddings — single provider mode
# -----------------------------------------------------------------------------
# EMBEDDING_PROVIDER one of:
#   local | openai_compatible | openrouter | groq | gemini | cloudflare |
#   cerebras | mistral | nvidia | puter
EMBEDDING_PROVIDER=local

# Local provider (sentence-transformers). Requires the "local" extra:
#   pip install sentence-transformers
EMBEDDING_MODEL=sentence-transformers/all-MiniLM-L6-v2

# Remote providers use an OpenAI-compatible /embeddings endpoint.
# EMBEDDING_API_KEY is the bearer token for the chosen provider.
# EMBEDDING_BASE_URL overrides the provider's built-in default base URL.
# EMBEDDING_API_KEY=
# EMBEDDING_BASE_URL=

# Example (Mistral remote embeddings):
#   EMBEDDING_PROVIDER=mistral
#   EMBEDDING_MODEL=mistral-embed
#   EMBEDDING_API_KEY=your-mistral-key
#
# Example (OpenRouter):
#   EMBEDDING_PROVIDER=openrouter
#   EMBEDDING_MODEL=openai/text-embedding-3-small
#   EMBEDDING_API_KEY=your-openrouter-key
#
# Example (Cloudflare Workers AI — base URL must include the account id):
#   EMBEDDING_PROVIDER=cloudflare
#   EMBEDDING_MODEL=@cf/baai/bge-base-en-v1.5
#   EMBEDDING_BASE_URL=https://api.cloudflare.com/client/v4/accounts/<ACCOUNT_ID>/ai/v1
#   EMBEDDING_API_KEY=your-cloudflare-token

# -----------------------------------------------------------------------------
# Embeddings — rotation mode (OPTIONAL)
# -----------------------------------------------------------------------------
# When configured, calls are spread across several providers/keys using a
# round-robin strategy with automatic failover on errors / 429 rate limits.
# All members MUST produce vectors of the same dimension.
#
# Strategy: round_robin (default) or random.
EMBEDDING_ROTATION_STRATEGY=round_robin
# Seconds a failed member is put on cooldown before being retried.
EMBEDDING_ROTATION_COOLDOWN=30

# Option A — JSON list (most explicit, recommended):
# EMBEDDING_ROTATION=[
#   {"provider":"groq","model":"text-embedding-3-small","api_key":"key1"},
#   {"provider":"mistral","model":"mistral-embed","api_key":"key2"},
#   {"provider":"openrouter","model":"openai/text-embedding-3-small","api_key":"key3"}
# ]

# Option B — parallel comma-separated lists (api_keys/models/base_urls are
# aligned by index with providers; shorter lists reuse the single-mode value):
# EMBEDDING_PROVIDERS=mistral,mistral,mistral
# EMBEDDING_API_KEYS=key1,key2,key3
# EMBEDDING_MODELS=mistral-embed
# EMBEDDING_BASE_URLS=

# -----------------------------------------------------------------------------
# MCP HTTP/SSE
# -----------------------------------------------------------------------------
# Path where the MCP SSE transport is mounted on the HTTP server.
MCP_SSE_PATH=/mcp