AI Input Enhancement (aiti-ai-input)

aiti-aiAbout 2 min

aiti-ai-input

❝ Before you finish typing, it has already guessed half of it. ❞

Input enhancement extension for the AI chat input box (long-term keeper, not a temporary experiment).

Autocomplete (L0 local instant → L1 AI)

Typing pauses 700ms with the cursor at the end → gray ghost text shows the next sentence (Tab/→ to accept, Esc/keep typing to cancel).

L0 local layer (zero-model, instant): first checks locally favorited prompts + session history inputs; the server POST /complete mode=local prefix-matches, and on a hit the gray text appears immediately with a ↦favorite/↦history badge, without calling the AI
L1 AI layer: only when L0 misses does it make a small same-model request (gray text appears after 0.4–2s)
Governance: extman feature management — autocomplete.enabled master switch · autocomplete.localOnly (L0 only, zero network egress) · autocomplete.minChars (minimum trigger length, default 3, CJK weighted ×1.5) · ctx three-tier context budget

Favorite suggestions (suggest)

Typing a prefix in the input box pops a candidate panel:

Static sources first: commands ⚡ rank0 + prompts 📌 rank1; when nothing matches, aiFallback serves as the fallback
Backend extension point: POST /providers {id,path,label,rank} registers a suggestion source
POST /suggest: aggregates provider results + strict filtering for short input (suggest.looseLen default 10)
⚡ one-click run: feature ai-input.suggestRun default off (enable in extman)

Voice input (three layers)

① Dictation detection + repair (default off: voice.fix / voice.indicator)

Pure browser event-feature detection of voice/high-speed input (beforeinput inputType + rhythm), zero server dependency
On detection → a "🎤 checking and fixing transcription errors for you" marker bar above the input box, dismissible with ✕
On send, AI spelling repair for typos/transliterated words/repeated words, giving replacement suggestions only at ambiguous spots

② 🎤 Real-time ASR (push to talk) (two master switches voice.enabled / toolbar.enabled, default off; enable in extman)

Real-time streaming WS recognition + VAD auto segmentation (~0.7s) + pseudo-real-time sentence segmentation via SenseVoice + on release a 2-pass offline refinement (Qwen3-ASR).

③ Pluggable ASR engines (asr-config.json)

Self-hosted 8767 (FunASR / Qwen3 / GLM): true streaming + model-level hotwords
iFlytek: local bridge 8768, true streaming
Baidu / SiliconFlow / Azure / OpenAI-compatible endpoints
Switch engines in the panel; keys are masked and only the last 4 characters are returned

Hotwords / usage stats / self-healing

Hotwords: the session-manager hotwords service auto-extracts per workspace and merges with the user lexicon (user-hotwords.json ≤500), pushing per-workspace biasing to 8767 recognition (hash key + 10min TTL debounce)
Usage stats: the "Voice traffic statistics" tab in the right panel (GET /asr/usage)
8767 self-healing: /asr/ensure launches it via the aiti-cmd runner PTY (circuit breaker 60s / dedup window 120s); /asr/control start|stop|restart