# codescout retrieval stack — CPU profile
#
# Source this before running codescout / sync_project:
# set -a; source .env.cpu; set +a
#
# Or pass to docker compose:
# docker compose --profile cpu --env-file .env.cpu up -d
# Compose profile selector (used by scripts/retrieval-stack.sh)
CODESCOUT_RETRIEVAL_PROFILE=cpu
# Host directory holding the GGUF model file(s) — bind-mounted into dense-cpu
# at /models. Download CodeRankEmbed-Q4_K_M.gguf into here before first run:
# mkdir -p ./models && \
# curl -L -o ./models/CodeRankEmbed-Q4_K_M.gguf \
# https://huggingface.co/brandtcormorant/CodeRankEmbed-Q4_K_M-GGUF/resolve/main/coderankembed-q4_k_m.gguf
CODESCOUT_MODEL_DIR=./models
# ---- Codescout client wiring ----
CODESCOUT_QDRANT_URL=http://127.0.0.1:6334
CODESCOUT_EMBEDDER_URL=http://127.0.0.1:48081
CODESCOUT_EMBEDDER_PROTOCOL=openai
CODESCOUT_EMBEDDER_MODEL_NAME=CodeRankEmbed-Q4_K_M.gguf
CODESCOUT_MODEL_DIM=768
# CodeRankEmbed is asymmetric: queries get this prefix, documents stay raw.
CODESCOUT_QUERY_PREFIX="Represent this query for searching relevant code: "
CODESCOUT_SPARSE_EMBEDDER_URL=http://127.0.0.1:48084
CODESCOUT_RERANKER_URL=http://127.0.0.1:48083
CODESCOUT_BM25_BOOST=3.0