Suggestive Replies

Admin access · 4-digit PIN

Rules

There's no model here. Every suggestion comes from a deterministic, auditable pipeline over plain keyword rules — nothing is sent off-device. The tradeoff is on purpose: it can't generalise like an LLM, but every match (and every non-match) traces back to a specific line in dataset.json, which is what you want when QA-ing hundreds of phrasings by hand. The whole thing rests on three ideas.

1 · Normalise before matching"HEYYYY!!", "heyyy" and "Hey" should all hit the same rule. Surface variation is stripped away first, so rules only describe intent — not every spelling of it.

2 · A keyword is a lead, not a verdict"We had a meeting" contains "meeting" but isn't scheduling one. "Call it a day" contains "call" without meaning a phone call. So matching is two phases: detect, then validate — a keyword hit can still be thrown out.

3 · Reject before generatingSuggesting the wrong thing is worse than suggesting nothing. Once a rule wins, it still has to survive a false-positive check before any reply is built.

Watch a message flow through

Pick an example or type your own. The funnel below runs the real matcher, stage by stage — the same code the app uses.

The pipeline, step by step

Normalise. Lowercase, strip punctuation, collapse elongated letters (byeee → bye), drop stopwords, lightly stem the rest (ing/ies/ed/s/ly), with a carve-out so boxes/catches aren't mangled.
Structural cases, checked first. A few signals skip keyword scoring entirely: a fenced ```code``` block, a URL, a short bare affirmation (sure, done...) matched as the whole message, and — only if nothing else matched — a bare time like "4 pm?". @mentions are detected separately and can attach an "Open X's chat" action to any message.
Score every rule. Each rule offers phrases (its trigger plus keywords). A phrase matches only if every one of its significant tokens appears — order doesn't matter. The most-specific matching phrase (most tokens) wins; ties go to whichever rule is defined first.
Validate the winner. Only the top rule is checked — against its exclude idiom list, and against recap detection (its own words in past tense, or "had a meeting"/"went to"). A live signal (?, "you", "let's", "can we") overrides the recap rejection.
Extract entities. Day (tomorrow...) and time (3 PM, half past 3, noon...) are pulled out and combined into {when} for templated replies.
Build the chips. Always exactly 3: the rule's action first (if any), then one reply per stance category (positive/negative/ask — never two from one), then a mood-matched filler tops up any gap. A reply keeps its baked-in emoji only if the incoming message used one.

The recurring gotcha: keyword collapse

Before adding any keyword phrase, run it through stopword removal and stemming in your head. "whats up" reduces to just ["what"] (up is a stopword) — as a keyword that matches any message with the word "what". Same trap: "works for you" → ["work"], "on leave" → ["leave"], "my bad" → ["bad"]. If a phrase collapses to one common word, find a safer multi-token phrasing, or add a targeted exclude regex for the bad pattern rather than dropping the coverage entirely.

What is fillerMood?

Every match returns exactly 3 chips. When a rule doesn't have enough of its own replies to fill all 3 slots, the gap is topped up with a generic filler — but one universal "Got it, thanks!" reads wrong everywhere. fillerMood just picks which pool of fillers to draw from, so the padding matches the rule's tone. Optional; defaults to neutral.

neutral	Reacting to being informed. e.g. "Noted, thanks." / "Sounds good."
warm	Friendlier acknowledgement. e.g. "Thanks for letting me know!"
casual	Light banter, nothing reported. e.g. "Ha, fair enough." / "Good stuff!"
empathetic	Bad news / sympathy. e.g. "Here if you need anything."
celebratory	Good news / wins. e.g. "That's fantastic!" / "Wonderful news!"
closure	Wrapping up. e.g. "Take care!" / "Have a great one!"

Dataset format

{
  "id": "rule_call",
  "trigger": "free for a call",
  "keywords": ["ring", "phone", "free to talk"],
  "exclude": ["call it a day", "\\bon call\\b"],
  "fillerMood": "neutral",
  "action": { "type": "call", "label": "Call" },
  "stances": {
    "positive": { "variants": ["Sure, ring me {when} 👍"] },
    "negative": { "variants": ["Can't talk right now, sorry."] },
    "ask":      { "variants": ["What's it about?"] }
  }
}

Action types: call, meeting, reminder, open-chat, share_slots, react. Only id, trigger and stances are required; everything else is optional. Slots {when}, {day} and {time} in a variant are filled from the extracted entities (or a safe default when the message had none).