The ranking model scores every listing and flags when a call is genuinely hard — the
top picks are a near-tie, or key specs are missing. When that happens (and this is on),
it asks a larger Escalation model for a second opinion. It's off by default
because the current ranking model is already the strongest one available, so escalation
mostly just adds time. Turn it on if you swap in a weaker/faster ranking model.
Web specs (SearXNG)
eBay listings for industrial parts often have almost no specs. With this on, the app
looks up each cable's model number on your self-hosted SearXNG search engine, grabs the
real datasheet numbers (impedance, loss, PIM, length), and hands them to the ranker — so
it judges on facts instead of guessing from the title. More accurate; adds a few seconds.
Intent / Ranking / Escalation model
Which model on your LiteLLM proxy handles each step. Intent turns your request
into a search; Ranking scores the listings; Escalation is the
second-opinion model. Leave on “(default)” to use the ones set in .env.
Max enrich
How many of the top listings get full specs pulled and ranked as cards (each costs one
eBay API call and one web lookup). Higher = more cards, more thorough, but slower.
Default 15. Large sets are ranked in batches automatically, so you can push this to 30-50
even on a smaller ranking model.
Import mode
Rank listings you paste in yourself instead of searching the eBay API — handy for
curating a specific set of listings by hand.