Install HUPI, set it up, and pick a tier.

Any Postgres works as long as the pgvector extension can be enabled on it — managed or self-hosted, nothing else about your setup is special-cased.

                    docker run -d --name hupi-pg \
  -e POSTGRES_PASSWORD=hupi \
  -e POSTGRES_DB=hupi \
  -p 5432:5432 \
  pgvector/pgvector:pg16
                  

The migrations in schema/ are numbered and must be applied in that order — later ones depend on tables and columns earlier ones create.

                    export HUPI_ADMIN_DATABASE_URL='postgres://postgres:hupi@localhost:5432/hupi?sslmode=disable'

for f in schema/0*.sql; do
  psql "$HUPI_ADMIN_DATABASE_URL" -f "$f"
done
                  

Name which AI vendor(s) to use for live chat, for consolidation, for the independent grounding check, and for embeddings — these can all be different vendors, or the same one. See providers.yaml.example.

HUPI_APP_DATABASE_URL should point at the restricted hupi_app role (not the admin role), not the superuser you used to apply migrations.

                    export HUPI_APP_DATABASE_URL='postgres://hupi_app:pick-a-real-secret@localhost:5432/hupi?sslmode=disable'
export HUPI_KEK=$(head -c32 /dev/urandom | base64)
export HUPI_PROVIDERS_CONFIG=./providers.yaml
                  

Listens on 127.0.0.1:8787 by default and speaks the standard OpenAI-compatible chat completions API — point any existing client at it.

                    ./bin/hupi
                  

Run hupi-consolidate once daily (cron) so conversations get distilled into searchable, fact-checked summaries overnight. Optionally hupi-selfcheck weekly once you have a few probe questions.