Parallel AI

Two scripts over the Parallel AI API. Prefer the built-in WebSearch tool for ordinary lookups, and the built-in deep-research skill for ordinary research — reach for these when you need domain-filtered cited excerpts, or a long report run outside this session.

Requirements

• PARALLEL_API_KEY — https://platform.parallel.ai

Search — cited excerpts

./scripts/parallel_search search -o "objective" [-q "keyword query"] [-d example.com] [-n 5]

JSON out: {objective, queries, mode, results:[{title, url, excerpt, publish_date}]}. Excerpts are markdown joined per URL; publish_date is often null.

Research — full reports

./scripts/parallel_research research "query" [-p core] [-t 1800] [-f markdown]

Processor cost per 1k runs: lite $5, base $10, core $25, core2x $50, pro $100, ultra $300, doubling to ultra8x $2400. Use core for routine multi-source questions and ultra only for genuinely hard ones. JSON out: {query, processor, run_id, content, basis}, where basis maps each output field to its source excerpts.

Gotchas

• search_queries is required by the API; omitting -q sends the whole objective as one query, which is worse than 2-3 tight keyword queries.
• Research runs 1-45 minutes and blocks until done — raise -t past the default 600s for ultra*, or the call times out while the run keeps billing.
• Query must be under 15,000 chars. Progress goes to stderr, results to stdout.
• The legacy -fast processor names (pro-fast, ultra-fast) still resolve but are not recommended for new work and cost the same as their standard counterparts.