# codescout retrieval stack — AMD ROCm profile
#
# Brings up qdrant + sparse-cpu + dense-amd + reranker-amd via codescout's own
# docker-compose. Dense and reranker run as `rocm/llama.cpp` containers with
# /dev/kfd + /dev/dri passthrough.
#
# Bring up:
# cd ~/work/claude/code-explorer
# docker compose --profile amd up -d
# Verify:
# curl 127.0.0.1:48081/health # dense (CodeRankEmbed on AMD)
# curl 127.0.0.1:48083/health # reranker (bge-reranker-v2-m3 on AMD)
# curl 127.0.0.1:48084/health # sparse SPLADE (CPU)
# curl 127.0.0.1:6333/healthz # qdrant
#
# Models needed in ${CODESCOUT_MODEL_DIR:-./models}:
# CodeRankEmbed-Q4_K_M.gguf (90 MB) — nomic-ai/CodeRankEmbed-GGUF
# bge-reranker-v2-m3-Q4_K_M.gguf (419 MB) — gpustack/bge-reranker-v2-m3-GGUF
#
# Source before running codescout:
# set -a; source .env.amd; set +a
CODESCOUT_RETRIEVAL_PROFILE=amd
# ---- Codescout client wiring ----
CODESCOUT_QDRANT_URL=http://127.0.0.1:6334
CODESCOUT_EMBEDDER_URL=http://127.0.0.1:48081
CODESCOUT_EMBEDDER_PROTOCOL=llama-server
CODESCOUT_EMBEDDER_MODEL_NAME=CodeRankEmbed-Q4_K_M.gguf
CODESCOUT_QUERY_PREFIX="Represent this query for searching relevant code: "
CODESCOUT_MODEL_DIM=768
CODESCOUT_SPARSE_EMBEDDER_URL=http://127.0.0.1:48084
CODESCOUT_BM25_BOOST=3.0
CODESCOUT_RERANKER_URL=http://127.0.0.1:48083
CODESCOUT_RERANKER_PROTOCOL=llama-server