# ─────────────────────────────────────────────────────────────────────
# nano-NOTEBOOKLM configuration
# Copy to `.env` and fill in the providers you actually use. You need
# AT LEAST ONE backend (openai / claude / local) configured for chat
# to work.
#
# After first boot, `artifacts/providers.json` becomes the source of
# truth — the Settings UI lets you add / edit / delete / test providers
# without restarting. The env vars below are the BOOTSTRAP SEED only:
# changing them after providers.json exists has no effect (operator
# edits win). Delete providers.json + restart to re-seed from env.
# ─────────────────────────────────────────────────────────────────────

# ─── 1. OpenAI-compatible provider (recommended) ─────────────────────
# Works with any service that exposes /v1/chat/completions — 16+ named
# vendors below plus OneAPI-style proxies and your own self-hosted
# gateways. Anything not in this cheat sheet still "just works" via
# the Settings UI as long as the endpoint is OpenAI-compatible.
# Cheat sheet — set OPENAI_BASE_URL + OPENAI_MODEL to one of:
#
#   OpenAI       https://api.openai.com/v1                          gpt-4o-mini
#   DeepSeek     https://api.deepseek.com/v1                        deepseek-v4-pro
#   Moonshot     https://api.moonshot.cn/v1                         moonshot-v1-8k
#   Zhipu        https://open.bigmodel.cn/api/paas/v4               glm-4-flash
#   MiniMax      https://api.minimax.chat/v1                        abab6.5-chat
#   Groq         https://api.groq.com/openai/v1                     llama-3.3-70b-versatile
#   Together     https://api.together.xyz/v1                        meta-llama/Llama-3.3-70B-Instruct-Turbo
#   Gemini       https://generativelanguage.googleapis.com/v1beta/openai/   gemini-2.0-flash
#   xAI Grok     https://api.x.ai/v1                                grok-4
#   OpenRouter   https://openrouter.ai/api/v1                       openai/gpt-4o-mini
#   Perplexity   https://api.perplexity.ai                          sonar-pro
#   Mistral      https://api.mistral.ai/v1                          mistral-large-latest
#   DashScope    https://dashscope-intl.aliyuncs.com/compatible-mode/v1   qwen-plus
#   Fireworks    https://api.fireworks.ai/inference/v1              accounts/fireworks/models/deepseek-v3p1
#   SiliconFlow  https://api.siliconflow.cn/v1                      Qwen/Qwen2.5-72B-Instruct
#   Cerebras     https://api.cerebras.ai/v1                         llama-3.3-70b
#
OPENAI_API_KEY=
OPENAI_BASE_URL=https://api.openai.com/v1
OPENAI_MODEL=gpt-4o-mini

# ─── 2. Anthropic Claude (native API, optional) ──────────────────────
ANTHROPIC_API_KEY=
CLAUDE_MODEL=claude-sonnet-4-5

# ─── 3. Local model via OpenAI-compatible /v1 (optional) ─────────────
# Ollama:    LOCAL_LLM_BASE_URL=http://localhost:11434/v1   model=qwen2.5:7b
# vLLM:      LOCAL_LLM_BASE_URL=http://localhost:8000/v1    model=<path-or-name>
# LM Studio: LOCAL_LLM_BASE_URL=http://localhost:1234/v1    model=<loaded model name>
# llama.cpp: LOCAL_LLM_BASE_URL=http://localhost:8080/v1    model=<model name>
LOCAL_LLM_BASE_URL=
LOCAL_LLM_MODEL=
LOCAL_LLM_API_KEY=local

# ─── Default backend ─────────────────────────────────────────────────
# Which one to use when a request doesn't specify. One of: openai, claude, local.
DEFAULT_BACKEND=openai

# ─── Embeddings ──────────────────────────────────────────────────────
# The Settings UI exposes three switchable presets (Local MiniLM /
# OpenAI text-embedding-3-large / BGE-M3). The chosen preset is
# persisted to `artifacts/embedding_preference.json` and OVERRIDES the
# env values below — env is only consulted as the first-run default.
#
# Each preset gets its own FAISS index namespace (`indices/faiss/<preset>/`)
# so toggling between them is a path-route, not a destructive rebuild;
# the first switch to a never-used preset triggers a one-shot background
# rebuild of every course's index.
#
# - "local" → sentence-transformers (offline, downloads on first use).
#           Default multilingual MiniLM handles 50+ languages incl. CJK.
# - "api"   → OpenAI-compatible /v1/embeddings. Falls back to OPENAI_* by
#           default; set EMBEDDING_API_BASE_URL / EMBEDDING_API_KEY to
#           route embeddings to a different provider than chat.
EMBEDDING_MODE=local
# EMBEDDING_MODEL=paraphrase-multilingual-MiniLM-L12-v2     # local_mini preset
# EMBEDDING_MODEL=text-embedding-3-large                    # openai_large preset
# EMBEDDING_MODEL=BAAI/bge-m3                               # bge_m3 preset
# EMBEDDING_API_KEY=
# EMBEDDING_API_BASE_URL=

# ─── Paths ───────────────────────────────────────────────────────────
# Where uploaded files, indices, KGs, notes, etc. live. Default is
# `./artifacts/` relative to the project root.
# ARTIFACTS_DIR=./artifacts

# ─── MinerU PDF extractor (optional) ─────────────────────────────────
# Install with: `uv pip install -e ".[mineru]"` (heavy: ~3-5GB w/ models).
# When engine=mineru is used for an upload, the backend keeps a
# persistent mineru-api subprocess alive so re-uploads don't pay the
# 30-45s model load each time. Tuning knobs:
#   MINERU_DEVICE_MODE — `cuda` / `cpu` / `mps`. Unset (default) → auto:
#       `cuda` when torch.cuda is available, else `cpu`. MPS is NOT
#       auto-selected (mineru pipeline hangs at DocAnalysis init on
#       Apple Silicon); set explicitly only if your mineru version
#       unblocks it.
#   MINERU_API_PORT — preferred port (default 47865); falls back to
#       an ephemeral port on collision.
#   MINERU_OMP_THREADS — caps OMP/MKL/OPENBLAS thread count inside the
#       mineru subprocess (default min(8, cpu_count)). Lower this when
#       running other compute alongside.
#   MINERU_API_MAX_CONCURRENT_REQUESTS — Linux only (Mac pins to 1).
#       Lets multiple PDFs run in parallel through the same server.
#   MINERU_SERVER_DISABLED=1 — force-fallback to the per-call subprocess
#       CLI (useful if the server is misbehaving in a deploy).
#   NANO_NLM_DISABLE_MINERU_WARMUP=1 — skip the boot-time warm-up; the
#       server still starts lazily on first request.

# ─── Retrieval / GraphRAG tuning (optional) ──────────────────────────
# All have sensible defaults; touch only if behavior needs adjustment.
#   GRAPHRAG_ENABLED=0 — kill switch for the KG-driven retriever. Set
#       to 0 to force every chat down the BM25 + vector RAG path
#       (useful when a course's KG is being rebuilt).
#   GRAPHRAG_SCORE_GATE_TOP1=0.15 — minimum top-1 cosine score for the
#       graphrag path to take over from RAG. Lower = more queries
#       routed to KG; higher = stricter admission.
#   NANO_NLM_DISABLE_EMBED_WARMUP=1 — skip boot-time embedding warm-up
#       (the model still loads lazily on the first request). Useful for
#       CI / containers where the warm-up wallclock hurts.