Per-URL extraction micro-benchmark. Fetches a URL once, runs the same
pipeline as --format llm, prints a small ASCII table comparing raw
HTML vs. llm output on tokens, bytes, and extraction time.
webclaw bench https://stripe.com # ASCII table
webclaw bench https://stripe.com --json # one-line JSON
webclaw bench https://stripe.com --facts FILE # adds fidelity row
The --facts file uses the same schema as benchmarks/facts.json (curated
visible-fact list per URL). URLs not in the file produce no fidelity
row, so an uncurated site doesn't show 0/0.
v1 uses an approximate tokenizer (chars/4 Latin, chars/2 when CJK
dominates). Off by ~10% vs cl100k_base but the signal — 'is the LLM
output 90% smaller than the raw HTML' — is order-of-magnitude, not
precise accounting. Output is labeled '~ tokens' so nobody mistakes
it for a real BPE count. Swapping in tiktoken-rs later is a one
function change; left out of v1 to avoid the 2 MB BPE-data binary
bloat for a feature most users will run a handful of times.
Implemented as a real clap subcommand (clap::Subcommand) rather than
yet another flag, with the existing flag-based flow falling through
when no subcommand is given. Existing 'webclaw <url> --format ...'
invocations work exactly as before. Lays the groundwork for future
subcommands without disrupting the legacy flat-flag UX.
12 new unit tests cover the tokenizer, formatters, host extraction,
and fact-matching. Verified end-to-end on example.com and tavily.com
(5/5 facts preserved at 93% token reduction).
