pmem-ai-semantic

Optional local semantic inference runtime for pmem-ai. It is intentionally distributed separately because Transformers.js includes large native inference and image-processing dependencies that deterministic pmem users do not need.

Install this package explicitly only on a machine where semantic retrieval is wanted:

npm install -g [email protected]
pmem semantic enable

pmem semantic enable performs the guided model setup and project index build. Operators can still run pmem semantic setup and pmem semantic rebuild separately.

The runtime loads the pinned model exclusively from the cache prepared by pmem semantic setup; remote model loading remains disabled during inference.

This companion intentionally contains the isolated Transformers.js native dependency graph. Its production dependency audit is run and reported separately from the base pmem-ai package so users who do not enable semantic retrieval never install that graph.

Repository contributors can opt in without changing the root manifest or lockfile:

npm install --no-save --package-lock=false ./packages/semantic-runtime
PMEM_RUN_SEMANTIC_EVAL=1 PMEM_SEMANTIC_CACHE="$HOME/.pmem-global/models/Xenova/multilingual-e5-small/<revision>" \
  node --require ts-node/register --test src/core/query/engine/semantic-evaluation.live.test.ts