Literess
Literess searches the web, and resources speak every language
From the Transept teamLiteress could always read your documents and reach for her tools, but nothing outside them. A question about a current product name, an unfamiliar acronym, or how a phrase is actually used in the wild got an answer from memory or no answer at all. She now searches the web as part of answering, and shows you what she read.
The rest of the release is about the resources that steer a translation. A glossary term and a style guide each hold a complete version per language now, instead of one entry stretched thin across all of them.
Added
- Literess searches the web. When a question turns on something she cannot know from your documents (a current fact, an idiom, a brand’s own spelling, a cultural reference), she searches for it while answering. A “Searching the web” pill shows the query as it runs, and the finished reply carries a numbered footer of the pages she used. It costs nothing: web search is part of her turn, not a run that spends words. Questions about Transept itself are answered from her built-in knowledge, not from the web. See What Literess can do.
- Long conversations compact instead of filling up. A row above the chat box shows how much of the conversation the model is still holding, with a capacity bar and the point where Transept steps in. Past that point it writes a summary of the conversation so far and carries on from it; you can also press Compact yourself at any time. Nothing is deleted: the older messages stay in the transcript behind a divider, and scrolling back still shows them.
- A complete glossary profile per language. Every extra language on a term now carries its own settings rather than one translation: the rendering (or a do-not-translate flag for that language alone), source aliases, other acceptable renderings, ones to avoid, gender and number, an example pair, and its own note. A language you have added but not filled stays empty everywhere: no other language’s value is shown or used in its place. The full term editor opens from the glossary page and from the settings icon on a term row in the editor sidebar. See Glossaries.
- Fill with AI, for the gaps only. A newly added language arrives empty. Fill with AI writes profiles for the terms that have none in that language and leaves every existing entry alone, so re-running after a partial fill costs only what is left. The run reads one of your documents (the open one in the editor, or one you pick from the glossary page), so the suggestions follow how the words are used there, drawing on each term’s occurrences, their current translations, and open comments on those passages. The dialog quotes an upper bound and bills only the profiles that land. Everything it writes arrives marked for review: usable in translations right away, flagged until someone approves it, with Approve and Reject on each entry and on the language as a whole. Editing one by hand approves it.
- Style guides hold a version per language. A language strip above the fields, samples and rules holds Language-agnostic plus a tab for each language you add. A language tab is a complete guide in its own right (nothing inherits, nothing merges), so a run translating into that language uses exactly that version, and a language with no version of its own uses the language-agnostic one. A guide that has no version for the document’s language says so on its sidebar card and offers Generate {language} version, which opens the run preconfigured. The result routes address form, quotation marks and number formats into that language’s version and leaves the shared voice alone, tells you so in the review, and changes nothing until you apply it. Reference samples come from real translations in the document where they exist. See Style guides.
- Literess’s edits appear in the document. When she proposes rewriting blocks, each one shows its before and after where the translation sits, with approve and reject beside it, and the rows on her confirmation card scroll the editor to the block in question. She can also turn the pages of the document she is reading (taking your view along or leaving it where it is), resolve comment threads she started herself once they are addressed, and tell you a reply has landed if you have wandered off. Proposals she has parked no longer expire after a day — they wait until you decide.
- Documents reopen where you left off. The editor remembers the page you were on and the block you were reading, per document, and returns you there the next time you open it. A link that names a page still wins, so a deep link you were sent goes exactly where it points.
- Add languages to an existing document over the API.
POST /public/v1/documents/:id/languagesgrows a language group that was created earlier, rather than forcing every language to be named up front; AI assistants get the same thing asdocument_add_languages. Projects can also be created inside a parent project now. See Automate Transept with the API.
Changed
- A style guide says which language version applies. The sidebar card names it (“Active · Ukrainian version applies” instead of the previous “Active · shared + Ukrainian”), and a chip per language lets you read any version’s resolved guidance without leaving the editor, with a note where a version does not apply to the document in front of you. “This doc” returns you to what runs actually read.
- Literess holds 200,000 tokens of conversation. Compaction fires as that fills, and the context row appears once a conversation is halfway through it.
- The team page hands its document list to the dashboard. The Documents card on a team page was a capped preview of a list that lives, filterable and complete, in the dashboard’s team view. It now shows the count and an Open in dashboard button that takes you there.
Fixed
- People search no longer returns strangers. Two characters typed into the add-people pickers matched accounts across all of Transept, names and email addresses included. A partial name or email now matches only people you already share a team or a document with, in either direction. A full email address still finds anyone, so inviting someone new works exactly as before, and both dialogs say so when a search comes back empty. A related gap on guest review links, which answered differently for an email that had an account than for one that did not, is closed: every email now gets the same answer, and people with accounts claim a link as themselves by signing in. See Teams and invites.
- Applying a generated style guide version keeps the guide attached. Applying one silently detached the guide from every document that had it active. Generated conventions could also land in the shared voice as English rules instead of in the language they were written for, and a guide whose own language could not be worked out could fail the merge outright. All three are fixed, and the review now shows the generated language version as its own labelled section.
- A deep link past the end of a document opens the last page. A page number carried over from a longer document, or a stale bookmark, rendered a perfectly healthy document as though it had no blocks. It now snaps to the real last page.
- The dashboard stops flipping between teams. Switching teams from the sidebar could set two updates chasing each other, and the view swapped teams until you navigated away.
Have a request or hit a snag? Ask Literess in the app, or write to [email protected].