.env.example

# ═══════════════════════════════════════════════════════════════
#  Tofu (豆腐) — Environment Configuration
# ═══════════════════════════════════════════════════════════════
#
#  Copy this file to .env and fill in your values:
#
#    cp .env.example .env
#
#  Then start the server:
#
#    python server.py
#
#  Priority: Settings UI > environment variables > defaults
#  (Any setting you configure in the web UI takes precedence)
#
#  Lines starting with # are comments. Remove # to enable a variable.
# ═══════════════════════════════════════════════════════════════


# ── LLM Provider ──────────────────────────────────────────────
# The easiest way is to configure via Settings UI (⚙️ → Providers).
# These env vars serve as fallback for headless / Docker deployments.

# API key (required — comma-separated for multiple keys)
# LLM_API_KEYS=sk-key1,sk-key2,sk-key3

# Single key (legacy, works too)
# LLM_API_KEY=sk-your-key-here

# API endpoint (default: https://api.openai.com/v1)
# LLM_BASE_URL=https://api.openai.com/v1

# Default model (default: gpt-4o)
# LLM_MODEL=gpt-4o

# Fallback model — used when the primary model fails (default: disabled)
# FALLBACK_MODEL=gpt-4o-mini


# ── Server ────────────────────────────────────────────────────

# Server port (default: 15000, auto-increments if occupied)
# PORT=15000

# Bind address (default: 0.0.0.0)
# BIND_HOST=0.0.0.0

# Flask debug mode (default: 0)
# FLASK_DEBUG=0


# ── Tunnel Authentication ─────────────────────────────────────
# Set to enable token-based auth for public tunnel access (e.g.
# VS Code port forwarding, ngrok). Leave empty for LAN-only mode.

# TUNNEL_TOKEN=your-secret-token


# ── Proxy (if behind a corporate firewall) ────────────────────

# HTTP_PROXY=http://proxy.example.com:8080
# HTTPS_PROXY=http://proxy.example.com:8080
# PROXY_BYPASS_DOMAINS=.internal.example.com,.corp.example.com


# ── Feishu / Lark Bot (optional) ──────────────────────────────
# Create an app at https://open.feishu.cn/app with Bot capability.

# FEISHU_APP_ID=cli_xxx
# FEISHU_APP_SECRET=xxx


# ── Feature Flags ─────────────────────────────────────────────

# Enable trading advisor module (default: 0 = off)
# TRADING_ENABLED=0

# Enable debug mode (default: 0 = off)
# DEBUG_MODE=0


# ── Search & Fetch ────────────────────────────────────────────
# These can also be configured in Settings UI (⚙️ → Search & Fetch).

# Number of search results to auto-fetch (default: 6)
# FETCH_TOP_N=6

# Per-page fetch timeout in seconds (default: 15)
# FETCH_TIMEOUT=15

# Max characters per page for search results (default: 60000)
# FETCH_MAX_CHARS_SEARCH=60000

# Max characters per page for direct URL fetch (default: 200000)
# FETCH_MAX_CHARS_DIRECT=200000


# ── PDF Parsing ───────────────────────────────────────────────

# Default text-extract strategy for /api/pdf/parse:
#   rich        — pymupdf4llm (default; ships out of the box)
#   structured  — IBM Docling (better tables + math on academic PDFs).
#                 Requires `pip install docling` (~2 GB; pulls torch).
#                 If docling is missing or fails, server falls back to
#                 rich automatically — uploads never break.
#   fast        — raw pymupdf get_text (no Markdown structure, ~50× faster)
# PDF_TEXT_MODE=rich

# VLM PDF parser tuning (applies when VLM mode is used).
# Pages per single VLM call (1–16). Default 4. Larger = fewer HTTP
# round-trips and less 429 thrash, but more output tokens per call.
# PDF_VLM_BATCH_PAGES=4

# Cap on concurrent VLM calls. Default = unlimited (one thread per
# batch). Lower this on shared keys to avoid 429 storms.
# PDF_VLM_MAX_WORKERS=8

# Output token cap per VLM call. Default scales with batch (4096/page).
# PDF_VLM_MAX_TOKENS=16384