.env.example

# Copy to .env and fill in. .env is git-ignored -- never commit real keys.
# These are read from the environment only; no key ever lives in source.

# ---- LLM provider (OpenAI-compatible; consumed ONLY by llm_client) ----
# Provider-agnostic on purpose: any OpenAI-compatible endpoint works with no code change.
# Never import or name a provider anywhere except llm_client.
#
# DEFAULT SINCE D-023: NVIDIA's hosted catalog (free NVIDIA Developer Program, no credit card).
# Get a key at https://build.nvidia.com
#
# NOT QWEN, AND THAT IS DELIBERATE. The judged submission answered with Qwen on Alibaba
# DashScope. No free Qwen text endpoint exists any more -- NVIDIA's hosted catalog carries
# no Qwen text model (verified against its own /v1/models: 102 models, zero Qwen), Groq's
# Qwen is paid-tier only, OpenRouter lists no free Qwen, and Cerebras serves none publicly.
# The answerer is not where this project's contribution lives: the zero-token WRITE path is
# local (surprise + NLI) and calls no API at all. Only the reply generator changed.
#
# The read path uses native function-calling, so keep to a model that supports tool use.
# List what your key can actually see:
#   curl -s -H "Authorization: Bearer $LLM_API_KEY" https://integrate.api.nvidia.com/v1/models
LLM_BASE_URL=https://integrate.api.nvidia.com/v1
LLM_API_KEY=
LLM_MODEL=meta/llama-3.3-70b-instruct

# Leave all three blank to run the dashboard offline: app.py falls back to MockLLM and
# labels its token counts as synthetic. The zero-token WRITE path is local either way.

# Judge model, used only by benchmarks/longmemeval.py -- not by the live app.
LLM_JUDGE_MODEL=

# ---- Embeddings (local, no API call) ----
EMBED_MODEL=BAAI/bge-small-en-v1.5

# ---- Compute device ----
# "cpu" everywhere by default; "cuda" only if you have a GPU build of torch.
DEVICE=cpu

# ---- OPTIONAL: Alibaba Cloud OSS snapshots (retired default, see D-022) ----
# Unset by default. Without these the app uses local-file snapshots and the
# cross-session boundary demo still works; only the backend label changes.
# Requires `pip install oss2`.
# ALIBABA_CLOUD_ACCESS_KEY_ID=
# ALIBABA_CLOUD_ACCESS_KEY_SECRET=
# OSS_ENDPOINT=https://oss-ap-southeast-1.aliyuncs.com
# OSS_BUCKET=