CanvasXpress MCP Server

# Environment variables for CanvasXpress MCP Server

# ============================================================
# LLM PROVIDER SELECTION
# ============================================================
# Choose your LLM provider: "openai" or "gemini"
LLM_PROVIDER=openai

# ============================================================
# AZURE OPENAI CONFIGURATION (when LLM_PROVIDER=openai)
# ============================================================
# Azure OpenAI API Key (from genai.web.bms.com)
AZURE_OPENAI_KEY=your_azure_openai_key_here

# Azure OpenAI API Version
AZURE_OPENAI_API_VERSION=2024-02-01

# Azure OpenAI Model (deployment name)
# Options: gpt-4o-mini-global, gpt-4o-global, gpt-4-turbo-global
LLM_MODEL=gpt-4o-mini-global

# BMS Environment (nonprod or prod)
LLM_ENVIRONMENT=nonprod

# ============================================================
# GOOGLE GEMINI CONFIGURATION (when LLM_PROVIDER=gemini)
# ============================================================
# Google AI API Key (from https://aistudio.google.com/apikey)
GOOGLE_API_KEY=your_google_api_key_here

# Gemini Model
# Options: gemini-2.0-flash-exp, gemini-1.5-flash, gemini-1.5-pro
GEMINI_MODEL=gemini-2.0-flash-exp

# ============================================================
# EMBEDDING PROVIDER SELECTION
# ============================================================
# Choose your embedding provider: "local", "onnx", "openai", or "gemini"
# - local: Uses BGE-M3 model locally (requires PyTorch ~2GB, proven 93% accuracy)
# - onnx: Uses lightweight ONNX models locally (no PyTorch, ~100MB, good accuracy)
# - openai: Uses Azure OpenAI text-embedding-3-small API
# - gemini: Uses Gemini text-embedding-004 API (faster startup, no local model)
EMBEDDING_PROVIDER=local

# ONNX embedding model (when EMBEDDING_PROVIDER=onnx)
# Lightweight local models (~1GB RAM vs ~3-4GB for BGE-M3 local)
#
# Recommended for light servers:
#   all-MiniLM-L6-v2      (384d, ~22MB, fastest)      <- DEFAULT
#   all-MiniLM-L12-v2     (384d, ~33MB, better quality)
#   multi-qa-MiniLM-L6-cos-v1 (384d, ~22MB, optimized for Q&A)
#
# Higher quality (more RAM):
#   all-mpnet-base-v2     (768d, ~420MB, best quality)
#   BAAI/bge-small-en-v1.5 (384d, ~33MB, BGE family)
#   BAAI/bge-base-en-v1.5  (768d, ~110MB, BGE family)
#
# Long context (8192 tokens):
#   nomic-ai/nomic-embed-text-v1.5 (768d, ~100MB, beats OpenAI ada-002)
#
# See: https://www.sbert.net/docs/pretrained_models.html
ONNX_EMBEDDING_MODEL=all-MiniLM-L6-v2

# OpenAI embedding model (when EMBEDDING_PROVIDER=openai)
# Options: text-embedding-3-small, text-embedding-3-large, text-embedding-ada-002
OPENAI_EMBEDDING_MODEL=text-embedding-3-small

# Gemini embedding model (when EMBEDDING_PROVIDER=gemini)
GEMINI_EMBEDDING_MODEL=text-embedding-004

# ============================================================
# MCP SERVER CONFIGURATION
# ============================================================
# Transport mode: http (network, default) or stdio (local, for Claude Desktop)
MCP_TRANSPORT=http

# HTTP mode settings (only used when MCP_TRANSPORT=http)
MCP_HOST=0.0.0.0
MCP_PORT=8000

# ============================================================
# PROMPT TEMPLATE CONFIGURATION
# ============================================================
# Prompt template version: v2 (default, includes rules) or v1 (original, no rules)
# v2 uses prompt_template_v2.md with canvasxpress_rules.md for better accuracy
# v1 uses the original prompt_template.md
PROMPT_VERSION=v2

# ============================================================
# ALTERNATIVE WORDING GENERATION (scripts/generate_alt_wordings.py)
# ============================================================
# Number of alternative wordings to generate per example
ALT_WORDING_COUNT=3

# Temperature for alternative wording generation (higher = more creative)
ALT_WORDING_TEMP=0.1