4 · The free verbs

Every command in this chapter runs at zero model calls. The habit to build: cut the corpus with free verbs first, then let the paid verbs judge only what is left.

Filter deterministically: where

cat logs.jsonl \
| smartpipe where 'level == "error" and not text contains "retry"' \
| smartpipe map "root cause? {cause}"

where is SQL-WHERE for your pipe: field predicates, comparisons, and/or/not. A typo in the predicate fails before reading stdin, with the operator menu.

Count and aggregate: summarize · chart · getschema

smartpipe getschema < tickets.jsonl                      # fields, types, coverage
smartpipe summarize 'count(), avg(total) by region' < tickets.jsonl
smartpipe where 'level == "error"' < logs.jsonl | smartpipe chart --by-time ts:1h

getschema first, always: it shows the dirt (mixed types, missing fields) before the dirt costs you skipped rows.

Trim and dedupe: sample · sort · distinct --exact · join --on

smartpipe sample 20 < corpus.jsonl        # seeded: same 20 every run
smartpipe distinct --exact < rows.jsonl   # fold byte-identical items, free
smartpipe join --on 'left.sku == right.sku' --right invoices.jsonl \
    --kind anti < orders.jsonl            # key-equality join, free

distinct without --exact and join with a prompt climb to embeddings and judge calls - the free forms exist so you only pay when meaning is actually needed.

While you iterate

smartpipe sample 20 on the front of a pipeline makes every experiment cost twenty items instead of the whole corpus. Combined with the result cache (chapter 5), iterating on a prompt costs almost nothing.

Next: 5 · Scale and cost - budgets, caches, and what happens when a provider goes down mid-run.