Echo Replies: the reply suggestion model

The model behind Settings → Local AI → Suggest replies. It reads a received message and offers a few short replies under it, worded in the message's language: a meeting request gets "That works for me." and "Let me check my calendar.", a German one "Das passt mir." and "Ich schaue in meinen Kalender.", news of a new job "Congratulations!". Choosing one starts a reply with it written in, ready to send or to go on writing.

It doesn't write text. It picks from the replies in catalog.json: 59 of them, each worded by hand in 37 languages (38 codes: Chinese in simplified and traditional characters). So every suggestion is a sentence a person wrote, and the model stays small and fast enough to run whenever a message opens.

File Used on Runtime
replies-v1.bin every platform Picacho's own (NativeTagger in src/manage/local_ai/runtime.rs)

Like the other models, it isn't kept in git: it is published on Hugging Face as pinaltech/echo-replies, and Picacho downloads it from there when the feature is turned on, checking its size and SHA-256 (REPLIES in src/manage/local_ai/mod.rs). A model trained again here with train.py needs a new file name, size and digest there, and publishing.

Languages

English, Spanish, German, French, Portuguese, Italian, Dutch, Swedish, Danish, Norwegian, Polish, Russian, Ukrainian, Czech, Slovak, Croatian, Bulgarian, Romanian, Hungarian, Finnish, Greek, Turkish, Arabic, Persian, Hebrew, Hindi, Bengali, Indonesian, Malay, Vietnamese, Thai, Japanese, Chinese (simplified and traditional), Korean, Tagalog, Swahili and Catalan: the date model's languages.

The wordings were written for this catalog and haven't been reviewed by native speakers yet. Where a language marks the speaker's gender in the verb (Hebrew, Hindi, Polish, Russian…), most wordings are phrased to avoid it; a few in Hindi ("मैं इसे देखता हूँ।", "मैं जाँच करके बताता हूँ।") are masculine. Review before relying on them.

How it works

  1. Cleaning. The body up to the sender's signature (--) or a forwarded or quoted original (-----, _____), without quoted lines (>), at most 4,000 characters.
  2. Words. The sign-in code model's (../otp/features.py): lowercase runs of letters, marks and digits, with scripts written without spaces split into pairs of characters.
  3. Features. The subject's words, the body's words and word pairs (the first 160), the words of up to four lines that ask something (?, ?, ؟), the concepts (lexicon.json) the body, those lines and the subject name, how many questions there are, whether the message exclaims, and its length. Concepts (invite, meet, call, thank, sorry, sick, late, cancel, deadline, birthday, holiday, introduce, interview, invoice, opinion, how-are-you, follow-up, news, loss, offer of help, automated…) list word stems in every language above, so "Einladung", "invitación" and "招待" all name invite. Everything is hashed into 32,768 slots.
  4. Network. The features, summed into 32 hidden units, then a softmax over the 59 replies and "none" (mail that wants no reply). The embedding is stored as 8-bit integers with a scale per row (the date model's PICAWHN1 format): about 1.2 MB.
  5. Suggestions (src/manage/local_ai/replies.rs). Nothing when "none" wins; else the likeliest replies (at most four, each at least 3%), worded in the message's language. The language is told from the script of its letters and, among languages sharing one, their common words (stopwords.json) and particular letters (ñ, ø, ő, ș…); when it can't be told, the computer's language is used, or English. What the person chooses makes it likelier (up to three times), and a reply turned away ("Don't Suggest This") isn't offered again; this history stays on the device and can be forgotten in Settings.

The app also offers nothing for mail from addresses like no-reply@, for mailing lists (List-Unsubscribe), sign-in codes, receipts and package updates, and for drafts and sent mail.

Data

languages.py and phrases_a.py to phrases_d.py write the mail; nothing in it is real. For each language there are, for each of 33 situations (thanks, a meeting request, an invitation, a request, a review, an attachment, news, bad news, a sick day, running late, an apology, a yes/no question, a call, a deadline, a check-in, wishes, an introduction, an interview, praise, a cancellation, a follow-up…) and automated mail, a few ways a message might say it, and greetings, small talk, sign-offs and subjects to wrap them in; SITUATIONS says which replies fit each. data.py builds messages of one or two situations, sometimes all lowercase, without question marks or accents, and trains on the share of each situation's fitting replies.

English has four wordings of each situation and the other languages two. That is the model's main weakness: real mail says things in many more ways. Training leaves out all the words of 30% of messages, keeping their concepts, so wording it never saw still names concepts it knows; more wordings, and native speakers' review of them, would help most.

Training

python3 train.py --quick   # a small run, scored on wording it never saw
python3 train.py           # numpy only: replies-v1.bin, classes.json, parity.json, eval.json

train.py first trains on half of the wordings and reports how the model does on the other half, by language, then trains the published model on all of them. Afterwards update REPLIES' size and SHA-256 in src/manage/local_ai/mod.rs. parity.json holds messages whose features, vectors and scores the app must reproduce (cargo test replies), and eval.json held-out messages in every language, on which the app's tests check that the language is told and the best suggestion fits.

catalog.json's wordings can change without retraining. Its list of replies can't: the model's outputs are in classes.json's order.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including pinaltech/echo-replies