[vidiyala99]

I ran my first parallel-cli findall run with the defaults (core, -n 10) on a new account. The published formula says that's at most $2.00 + 10 x $0.15 = $3.50. The run was billed $3.75, and the Usage page showed "Task API v1: 11" on a day I never called the Task API.

The difference is enrichments. run_findall applies whatever enrichments ingest suggests (enrich=True), each one billed per match at its Task processor rate. For my objective ingest added a founding_date enrichment on core, so 10 x $0.025 = $0.25. The CLI never mentions this, and there's no way to turn it off. --dry-run doesn't help either, because ingest runs again when the real run starts and isn't deterministic: the same objective came back with 0, 1, then 0 suggested enrichments across three calls.

At -n 10 it's a quarter. At -n 1000 with a couple of core enrichments it's an extra $50 nobody asked for.

What this changes, all in findall run:

• As soon as ingest returns, it prints the enrichments the run will add (field names and processor) and an estimated max cost, e.g. Est. cost: up to $3.75 ($2.00 + 10 matches x $0.175). The same estimate is in --dry-run, in both text and --json (estimated_cost, apply_enrichments).
• --no-enrich skips suggested enrichments.
• -g preview is back for findall run. The docs say to "Always run with generator: "preview" first", but the CLI only accepted base|core|pro. Preview was removed in feat: add FindAll to DuckDB integration + new CLI commands #58 because it doesn't support enrichment reliably, so it stays out of FINDALL_GENERATORS (the DuckDB paths are unchanged), enrichment is always skipped for it, and --match-limit must be 5-10.
• Fixes --dry-run printing enrichments as - : . It read name/description keys that ingest enrichments don't have; it now lists the fields and processor.

The prices come from https://docs.parallel.ai/getting-started/pricing and live in FINDALL_GENERATOR_PRICING and TASK_PROCESSOR_PRICE_PER_RUN in core/findall.py, with estimate_findall_cost() and summarize_enrichments() alongside. An unknown generator or enrichment processor makes the estimate say "unknown" rather than guess. I know hardcoded prices can drift, so if there's an API or a better source for these I'm happy to switch to it. That would also help with #158 and parallel-sdk-python#17.

Live run with this branch:

$ parallel-cli findall run "Find AI web search API companies founded after 2022" -g preview -n 5
Starting FindAll with generator=preview, match_limit=5
Generator: 5-10 candidates - test a query cheaply before a full run

Schema ready - entity type: companies
  Enrichments skipped: 1
    - founding_date (processor: core, billed per match)
  Est. cost: up to $0.10 ($0.10 + 5 matches x $0.00)
...
Status: completed | generated: 5, matched: 3 (1m52s)

The same objective on the defaults took 5m20s and $3.75.

Testing: 25 new tests in tests/test_findall.py (cost estimates including the $3.75 case above, run_findall with enrich=False and preview, and the CLI flag, dry-run and validation paths). ruff check, ruff format --check, tach check and ty check are clean. Full pytest: 864 passed, 14 failed, and the same 14 fail on main on my Windows machine (Unix permission, symlink and DuckDB event loop tests), so they're unrelated.

I used AI assistance while working on this and reviewed and tested the change myself.

🤖 Generated with Claude Code