What’s new in VelvetSpeak v1.2.1

VelvetSpeak v1.2.1 is on its way, and the biggest news since our v0.9.7 update is the suggestion-engine upgrade in v1.2.0. We’ve reworked how Velvet uses conversation context, builds longer replies, draws on preparation, and delivers answers to your glasses. v1.2.1 adds a better way to tell us how those suggestions work for you.

The main update: v1.2.0’s suggestion engine

In our last post, we said suggestions needed to be more relevant and natural, and fact-checking needed to be tighter. Much of the work since then has gone into the engine behind those moments.

More complete answers, with room to develop a thought. Building on the connected-answer work in v1.1.0, the upgraded engine treats a multi-page reply as one answer you can follow in order. Medium and Long are directed to develop the reasoning, relevant examples and a natural ending. v1.2.0 adds more display room for Long answers and checks that the complete response survives pagination, addressing answers being squeezed into an overly small card limit. Short remains focused on a direct reply.

A stronger grasp of who said what. The engine now gets clearer distinctions between your words, the other person’s words, and speech whose speaker is uncertain. Its instructions preserve who owns a fact, preference or commitment, and give newer corrections priority for that same person. This targets mistakes such as borrowing someone else’s experience for your answer or treating a proposed deadline as a promise you’ve already made.

Better handling of conversation history. We’ve revised how recent speech is retained when context fills up. The engine receives clear signals when earlier material is missing, and retained excerpts preserve complete sentence units. Recalibrating within a session also preserves earlier speech history. The aim is to keep useful context without quietly changing its meaning.

Preparation that carries through into the answer. The upgraded context path gives selected Plan fields and key facts a clearer place in the request, with more room in the expanded preparation mode. Private Plan Notes are excluded from live-answer context. We’ve also built a résumé review flow: upload or paste a document, review the extracted text, then explicitly choose whether to use it for Interview answers. Saving a document and activating it are separate choices, and it can be withdrawn.

More useful follow-up answers. “Another” and “Explain” carry the selected answer forward alongside the latest conversation. Another is directed to offer a meaningfully different move; Explain develops the reasoning and adjusts to newer information. Earlier generated advice is kept separate from things you actually said, so an invented detail should not become a supposed fact in the next reply.

Research that can inform a spoken response. For supported requests in Interview, Sales, Meeting and Debate, the new research path can feed requested web-search evidence into answer generation. It keeps source information available separately and makes unavailable evidence explicit, with instructions against implying that an unsupported claim has been verified.

An earlier start to reading. We’ve added an early-card delivery path for manual answers, designed to let you begin reading complete opening cards while the rest is generated. It protects an answer you’re already reading from being replaced and marks an interrupted partial result as incomplete.

These engine capabilities have been implemented with separate rollout controls. Advanced features are being validated through controlled testing before wider enablement; installing the latest app does not turn every capability on at once. Résumé-based live answers remain switched off during validation. These changes are aimed at better relevance, depth and delivery—we still need real-conversation feedback to judge how well they achieve that.

What v1.2.1 adds

The latest patch focuses on feedback. After a session, open Send feedback on your phone, choose a suggestion, and tell us whether it helped or missed the mark: too generic, too late, the wrong facts or person, the wrong length, or simply not your voice.

You can optionally attach a previewed suggestion excerpt and, separately, a short conversation excerpt. Both sharing boxes start unchecked, and you can send feedback without either attachment.

There is also an updated consent prompt explaining account-linked measurements of answer openings, page views and reported glasses delivery. Those measurements contain no conversation or suggestion text and do not establish that you spoke a suggestion aloud. Excerpts are sent only when you choose to attach them to a support report; they are not collected automatically or used automatically for model training. Measurements and submitted reports have no automatic expiry and remain until account deletion or operator cleanup.

Other improvements since v0.9.7

You can try a fictional sample conversation on your phone without using the microphone or saving a conversation; each try uses one answer from your allowance. We’ve also tightened speech-connection handling during reconnects and made turning off the glasses transcript turn off Draft consistently.

When v1.2.1 becomes available to you, tell us about one suggestion that helped—or one that should have. The engine work is about making Velvet more useful in the moment, and those specific examples help us see what still needs work.