Class: Collavre::TypoCorrectionsController

Inherits:
ApplicationController show all
Defined in:
app/controllers/collavre/typo_corrections_controller.rb

Overview

Session-authenticated endpoint backing the inline typo-correction UI. The frontend also gates by device/location, but we re-check here (fail closed) before spending an LLM call, and echo the user’s auto-apply threshold so the client never has to hardcode it.

Constant Summary collapse

MAX_TEXT_LENGTH =

The client debounce is only a hint — a scripted/compromised page can loop this endpoint with arbitrary text, and every accepted call spends an LLM request. Bound both the call rate and the per-call size server-side. The chat composer is short text; anything larger is not a typo-correction case.

4_000
RATE_LIMIT =

requests

120
RATE_PERIOD =

…per user per minute

1.minute

Instance Method Summary collapse

Instance Method Details

#create ⇒ Object 



15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
 
# File 'app/controllers/collavre/typo_corrections_controller.rb', line 15

def create
  user = Current.user
  device = params[:device].to_s
  location = params[:location].presence || "chat"
  text = params[:text].to_s

  unless user.typo_correction_active_for?(device: device, location: location)
    return render json: { edits: [], threshold: user.typo_correction_threshold }
  end

  # Oversized input: skip the LLM entirely (return no edits, not an error —
  # this is a background affordance, not a user-facing failure).
  if text.length > MAX_TEXT_LENGTH
    return render json: { edits: [], threshold: user.typo_correction_threshold }
  end

  # Raises RateLimitExceeded -> 429 (DynamicRateLimiting), which the client
  # treats as "no suggestions this round".
  check_rate_limit!(key: "typo_correction:#{user.id}", limit: RATE_LIMIT, period: RATE_PERIOD)

  edits = TypoCorrector.new.correct(text)
  render json: { edits: edits, threshold: user.typo_correction_threshold }
end