Recipe: Route Decisions Without Row Loops

When throughput matters, do not loop over every row and inspect a Python string one by one. Ask BatchResult for row groups and pass those row indices to your downstream writers.

result = engine.evaluate_ndjson(payload)

approved = result.indices_for_decision("APPROVE")
blocked = result.indices_for_decision("BLOCK")
needs_action = result.indices_for_not_decision("APPROVE")
all_groups = result.grouped_decision_indices()

CLI equivalent: emit decision rows instead of routing in Python

cat events.ndjson | \
  blazerules_agent \
    --rules rules.yaml \
    --input stdin \
    --output ndjson \
    --output-path decisions.ndjson

Tail or consume decisions.ndjson and route by the decision field. For high-throughput Python routing, use the grouped index arrays shown above.

Use these arrays to slice your original Arrow batch, Kafka metadata list, or application buffer.

def send_rows(writer, rows):
    if len(rows):
        writer.write_indices(rows)

groups = result.grouped_decision_indices()
send_rows(service_a, groups.get("APPROVE", []))
send_rows(service_b, result.indices_for_not_decision("APPROVE"))

CLI equivalent: split compact decision logs by decision

jq -c 'select(.decision == "APPROVE")' decisions.ndjson > approve.ndjson
jq -c 'select(.decision != "APPROVE")' decisions.ndjson > action.ndjson

If you need rule-level routing, use winning-rule groups:

by_rule = result.grouped_winning_rule_indices()
high_amount_rows = by_rule.get("high_amount_foreign_or_rooted", [])

CLI equivalent: split by winning rule from decision logs

jq -c 'select(.winning_rule_id == "high_amount_foreign_or_rooted")' \
  decisions.ndjson > high_amount_rows.ndjson

For detailed debugging, enable OutputDetail.BITMASKS and use:

rows = result.indices_for_rule("high_amount_foreign_or_rooted")
mask = result["high_amount_foreign_or_rooted"]

BITMASKS is useful for audits and rule debugging, but DECISIONS is the cheaper production default for routing.