.env.example

# ---------------------------------------------------------------------------
# Fully on-box, no hosted API. The answer model is a local OpenAI-compatible
# server (llama.cpp serving Qwen3.5-9B) and embeddings come from a local
# embedding server (llama.cpp serving bge-large-en-v1.5).
LLM_BASE_URL=http://127.0.0.1:8400/v1
LLM_API_KEY=local
EMBEDDING_BASE_URL=http://127.0.0.1:8401/v1
DISABLE_THINKING=1
# With DISABLE_THINKING=1 (the default) the model emits only the visible answer;
# if thinking is enabled the answer cap must also cover the reasoning trace.
ANSWER_MAX_TOKENS=2048
SUMMARY_MAX_TOKENS=512
# bge-large-en-v1.5 tops out at 512 tokens, so chunks are kept at 256/32
# (400/50 would overflow the embedder's tokenizer past 512).
CHUNK_TOKENS=256
CHUNK_OVERLAP=32
# ---------------------------------------------------------------------------

OPENAI_MODEL=qwen3.5-9b-local
OPENAI_EMBEDDING_MODEL=bge-large-en-v1.5
EMBEDDING_DIMENSIONS=1024
OPENAI_EMBEDDING_TIMEOUT_S=45
OPENAI_EMBEDDING_MAX_RETRIES=0
PARALLEL_RAW_RETRIEVAL=1
SETTLED_DRAFT_DELAY_MS=500
TRIGGER_TIMEOUT_S=4.0
RETRIEVAL_TIMEOUT_S=6.0
ANSWER_TIMEOUT_S=30.0
SUMMARY_TIMEOUT_S=8.0
POST_ANSWER_PERSISTENCE_TIMEOUT_S=10.0
TURN_IDLE_TIMEOUT_S=120
SESSION_RETENTION_HOURS=24

# Empty URL uses embedded local Qdrant. Set these only when running against a
# separately managed local Qdrant server.
QDRANT_URL=
QDRANT_API_KEY=

# Standalone Naive and Stream services default to separate ./var/<role>/ paths.
# Uncomment these only when intentionally overriding one service's state.
# QDRANT_PATH=./var/custom/qdrant
# RUNTIME_DB=./var/custom/runtime.sqlite3
# METRICS_LOG=./var/custom/requests.jsonl

# The single evaluation dataset is intentionally review-gated. Set to 1 only for local
# implementation checks before human approval; never label those runs final.
ALLOW_UNREVIEWED_DATASET=0

CORS_ORIGINS=http://localhost:5173,http://127.0.0.1:5173