render.yaml

# Render blueprint — one-click deploy of the Text-to-SQL API.
#
# Render reads this on "New → Blueprint" and configures the service, so nothing
# has to be typed into a form except the secrets.
#
# The database is NOT defined here. Render's free Postgres expires after 90
# days; Neon's free tier does not, and this database is only 61 MB. So Postgres
# lives on Neon and its credentials arrive as environment variables.

services:
  - type: web
    name: text2sql-api
    runtime: python
    plan: free
    region: oregon
    branch: master

    buildCommand: pip install --no-cache-dir -r requirements.txt
    # $PORT is assigned by Render at runtime; binding to a fixed port fails.
    startCommand: uvicorn src.api.main:app --host 0.0.0.0 --port $PORT

    # Hits the real readiness endpoint, which executes SELECT 1 against
    # PostgreSQL. A process that is listening but cannot reach its database is
    # not healthy, and a TCP check would call it healthy.
    healthCheckPath: /health

    envVars:
      - key: PYTHON_VERSION
        value: "3.11.9"
      - key: PYTHONPATH
        value: "."

      # Serves the BASE model. The fine-tuned adapter needs a GPU, which the
      # free tier does not have, so this is NOT the 70.86 % pipeline.
      - key: MODEL_BACKEND
        value: hf
      # Prompt v2 adds a business glossary and a DATA_AS_OF reference date.
      # Costs nothing, needs no GPU, and is worth +9.47 pp to the base model
      # (34.22 % -> 43.71 % strict, both scored on benchmark v2).
      - key: PROMPT_VERSION
        value: v2
      # Failover chain, tried in order. Free providers flip between fine,
      # busy, timing out and out-of-credits; depending on one took the demo
      # down twice. featherless-ai leads because on 2026-09-23 nscale was
      # returning 504 only after a 121 s timeout, hanging every request.
      - key: HF_PROVIDER
        value: featherless-ai,nscale
      # A person is waiting, so a dead provider is abandoned in seconds rather
      # than the 120 s a benchmark run can afford.
      - key: HF_TIMEOUT_S
        value: "30"
      - key: HF_MAX_RETRIES
        value: "4"

      # Managed Postgres refuses plaintext connections.
      - key: PGSSLMODE
        value: require
      - key: PGPORT
        value: "5432"

      - key: SQL_STATEMENT_TIMEOUT_MS
        value: "30000"
      # Neon's free tier caps concurrent connections; a small pool leaves room
      # for the seeding script and psql to connect alongside the API.
      - key: DB_POOL_MAX
        value: "3"
      - key: DB_POOL_MIN
        value: "1"

      # sync: false means Render prompts for these instead of storing them in
      # this file. Nothing secret is ever committed.
      - key: PGHOST
        sync: false
      - key: PGDATABASE
        sync: false
      - key: APP_DB_USER
        sync: false
      - key: APP_DB_PASSWORD
        sync: false
      - key: HF_TOKEN
        sync: false