# codescout retrieval stack — LITE profile (daemon-free)
#
# For a locked-down VDI or air-gapped box that CANNOT run Docker or Qdrant.
# Code search + memory run in-process on sqlite-vec (statically-linked vec0 —
# no foreign DLL, so an EDR like CrowdStrike has nothing to quarantine), with
# dense embeddings from a remote OpenAI-compatible endpoint. No sparse server,
# no reranker, no Qdrant. See docs/plans/2026-06-16-two-stack-retrieval-lite.md.
#
# Source before running codescout:
# set -a; source .env.lite; set +a
# Select the in-process sqlite-vec backend (this is what makes it "lite").
CODESCOUT_VECTOR_BACKEND=sqlite-vec
# Where the per-project sqlite-vec DBs live.
# Default: <home>/.codescout/embeddings
# CODESCOUT_SQLITE_DIR=/path/to/embeddings
# Remote OpenAI-compatible embedding endpoint — corporate gateway, OpenAI,
# Ollama, vLLM, or a local llama-server. Dense is always OpenAI-compatible:
# codescout POSTs {url}/v1/embeddings.
CODESCOUT_EMBEDDER_URL=https://embed.corp.example/v1
CODESCOUT_EMBEDDER_MODEL_NAME=your-model-name
CODESCOUT_MODEL_DIM=768
# Bearer token for an authenticated endpoint. HTTPS strongly recommended when set.
# EMBED_API_KEY=...
# Some asymmetric models want a query-side prefix (left empty by default):
# CODESCOUT_QUERY_PREFIX="Represent this query for searching relevant code: "
# No Qdrant / sparse / reranker URLs are needed — the lite stack ignores them.