Labs

PLAYBOOK SUMMARY

Keep it simple and consistent:

Keys: clear, user-shaped:
what is X, how does X work, what materials are used in X, can X be used for Y, benefits of X, etc.

First 200 chars of the answer: include the must-have tokens the user is likely to type. That’s what our content signal reads first.

Avoid duplicate keys; when two topics are close, make the keys explicitly distinct (e.g., add the disambiguator word).

If a crucial term only appears in the answer (like neodymium), put it early once—our prescan/nudge now catches it reliably.

What each knob does

kw (key weight): how much to trust the question-to-key match. Bigger kw = key dominates.

cw (content weight): how much to trust the first ~200 chars of the answer. Bigger cw = preview content can move results more.

kg (gate): if the key match is below this number, content is ignored. (Prevents weak keys from being dragged by content.)

floor (content floor): if content score is below this number, treat content as 0. (Filters noisy previews.)

cap (content cap): even when content helps, final score can’t exceed key + cap. (Stops runaway boosts.)

“key=100?”

That’s just the display scale in debug (0–100). For short-circuited near-exact matches we still keep the raw 990/1000 internally and also show key_raw=990 in the debug line you added. So:

• key_used ≈ the 0–100 value used in the math (easy to read)

• key_raw = the original 990/1000 you care about for exact/near-exact detection

Labs usage crystal clear (tiny tweak)

Flip Labs form with “Send overrides (tuning)” checkbox. When it’s unchecked, the script doesn’t include kw/cw/kg/cap/floor in the JSON—so you’re guaranteed to mirror production. When it’s checked, it includes them—so you’re in tuning mode.

CONSOLE TEST

fetch(“https://ak1amc.pythonanywhere.com/debug_match”,{ method:”POST”, headers:{‘Content-Type’:’application/json’}, body:JSON.stringify({ question:”neodymium“, topic:”general“, kw:0.60, cw:0.40, kg:20, cap:20, floor:35 }) }).then(r=>r.json()).then(console.log)