sources.example.yaml

# ingestlib structured-retrieval sources — query your databases alongside
# your documents via retrieve(sources=[...]). Optional: create sources.yaml
# only to use structured retrieval; it lives beside config.yaml, like
# rules.yaml. Connection URLs are secrets — set them in .env and reference
# them here as ${VAR}. Each database needs its pip extra, e.g.
# pip install "ingestlib[postgres]".
#
# Everything below is a commented example — there are no default sources.
# Uncomment and edit for your own databases.

# ── A SQL database — a connection plus a permission boundary ─────────────
# The model GENERATES read-only SQL inside this boundary; nothing can write —
# the DB role is read-only and `allow` caps which statement types may run.
# prescriptions:
#   type: postgres                  # postgres | mysql | sqlite | duckdb | snowflake
#   dsn: ${RX_DB_DSN}               # a READ-ONLY connection URL, set in .env
#   description: "Prescription fulfillment — status, ready times, patients, drugs, refills."
#   allow: [select]                 # statement types the model may generate (default: read-only)
#   row_limit: 1000                 # cap rows returned
#   timeout: 10                     # seconds before a query is killed
#   tables:                         # optional schema hints — the accuracy lever
#     rx: "one row per prescription — rx_id, patient_id, drug_name, status, ready_at, refills_left"
#     patients: "prescription owners — patient_id, name, dob, plan_id"
#   # Wide schema? Retrieve only the tables a question needs (plus their FK
#   # bridges) instead of dumping every table into the prompt. auto (default)
#   # dumps small schemas and retrieves on large ones; on = always retrieve;
#   # off = always dump. Generate hints for a cryptic schema with
#   # `ingestlib describe-schema <name>`; measure accuracy with `ingestlib eval-sql <name>`.
#   schema_rag: auto                # auto | on | off
#   schema_rag_top_k: 15            # tables retrieved per question (before FK closure)
#   schema_rag_min_tables: 10       # auto: dump all at or below this table count
#   # Optional accuracy override — a matching question runs YOUR reviewed SQL
#   # instead of a generated one, for answers that must be exact:
#   verified:
#     rx_status:
#       description: "Fulfillment status and ready time for a prescription"
#       sql: "SELECT status, ready_at FROM rx WHERE rx_id = :rx_id"
#       params: [rx_id]

# ── A document source — your existing ingestlib corpus ───────────────────
# package_inserts:
#   type: documents                 # reuses parse → chunk → embed → retrieve
#   namespace: package_inserts      # which corpus partition to search