.env.example

# mindex docker-compose tunables.
#
# Copy to a profile (e.g. perf/env/baseline.env) or to .env, then:
#   docker compose --env-file perf/env/baseline.env up -d --force-recreate mindex
# A plain .env in this directory is loaded automatically by `docker compose up`.
#
# These set ONLY the mindex container's flags. The embedder runs on the host and is
# NOT in compose — change its serving knobs (llama.cpp's -np/--ubatch-size, vLLM's
# --gpu-memory-utilization) wherever that server is configured, not here.

# Chunks per /v1/embeddings call — the GPU-batch lever. The server does its own
# batching, so this sizes the request rather than a forward pass.
MINDEX_EMBED_BATCH=1024

# SQLite connection pool. Indexer/k6 concurrency must stay <= this or requests get
# 500 (PoolEmpty); leave headroom for the GC/retry workers.
MINDEX_DB_POOL_SIZE=32

# Max /index request body (MiB). The name carries the unit, like every other
# sized key — and this one spelled it `_MB` for three releases while
# docker-compose.yml read `_MIB`, so setting it here was silently ignored and the
# compiled default took over. An over-cap request is a 413 `request.body_too_large`,
# which reads as a client that sent too much rather than a knob that never applied.
MINDEX_MAX_BODY_MIB=256

# Minutes a file may sit in `indexing` before the retry worker treats it as stuck.
# Must exceed the longest legitimate in-flight batch.
MINDEX_STUCK_GRACE_MINS=30

# Where mindex reaches the embedder (host-run; see deploy/embedder/). 11211 is
# what its unit and its recipes use; llama-swap would be 12434.
MINDEX_MODEL_SERVER=http://host.docker.internal:11211

# Log level.
MINDEX_RUST_LOG=info