Per-meeting language selection for AI summaries

[hyper-charlie]

Feature Request

Currently, AI-generated summaries are produced in the language set in the app's language settings. For users who attend meetings in multiple languages, this creates a mismatch—a meeting conducted in English produces a Polish summary (or vice versa).

Desired Behavior

1. Manual language selection: Allow users to select the summary language per meeting (e.g., from a dropdown or language picker in the meeting notes view)
2. Automatic language detection (optional enhancement): Automatically detect the meeting's spoken language and generate summaries in that language

Use Case

A user attends meetings in both English and Polish. Currently, with pl-PL locale settings, all summaries are generated in Polish regardless of the meeting's spoken language. This requires manual translation or re-summarization.

Environment

• Platform: macOS
• App Version: 1.0.17
• User Locale: pl-PL

[devin-ai-integration]

Codebase Exploration: Per-Meeting Language Selection for AI Summaries

How language currently flows through the system

The summary language is determined by a single global setting (ai_language) with no per-session override. Here's the full data flow:

1. Settings UI — apps/desktop/src/settings/general/main-language.tsx renders a MainLanguageView dropdown under "Language & Region". It reads/writes the ai_language value in the TinyBase settings store.

2. Settings persistence — apps/desktop/src/store/tinybase/persister/settings/persister.ts loads the setting from disk. If unset, it falls back to the OS locale via detectCommands.getPreferredLanguages() (from @hypr/plugin-detect). This is why a pl-PL locale user gets Polish summaries by default.

3. Enhance transform — When a summary is triggered, apps/desktop/src/store/zustand/ai-task/task-configs/enhance-transform.ts (line 96–99) reads the global setting:

   function getLanguage(settingsStore: SettingsStore): string | null {
     const value = settingsStore.getValue("ai_language");
     return typeof value === "string" && value.length > 0 ? value : null;
   }

   This is the only place the language is resolved for enhancement — there is no per-session or per-transcript override.

4. System prompt — The language string is passed into the EnhanceSystem struct (crates/template-app/src/enhance.rs) and rendered via the Jinja template at crates/template-app/assets/enhance.system.md.jinja:

   You are an expert at creating structured, comprehensive meeting summaries in {{ language | language }}.
   

   The language filter (crates/askama-utils/src/filters.rs) converts the ISO 639 code to a human-readable name (e.g., "pl" → "Polish"). It also conditionally adds non-English guidelines (keep technical terms in English, Korean-specific formatting, etc.).

5. Title generation follows the same pattern — title-transform.ts also reads settingsStore.getValue("ai_language").

Relevant files

File	Role
apps/desktop/src/store/zustand/ai-task/task-configs/enhance-transform.ts	getLanguage() — resolves summary language from global settings
apps/desktop/src/store/zustand/ai-task/task-configs/enhance-workflow.ts	getSystemPrompt() — passes language into template rendering
crates/template-app/assets/enhance.system.md.jinja	System prompt template that uses the language
crates/template-app/src/enhance.rs	EnhanceSystem struct (holds language: Option<String>)
crates/askama-utils/src/filters.rs	language, is_english, is_korean filters
apps/desktop/src/settings/general/main-language.tsx	Global "Main language" dropdown UI
apps/desktop/src/store/tinybase/persister/settings/persister.ts	Settings loading + OS locale fallback
apps/desktop/src/services/enhancer/index.ts	EnhancerService.enhance() — orchestrates the enhancement flow
crates/language/src/detect.rs	Text-based language detection (using whichlang) — currently unused for summaries
packages/store/src/tinybase.ts (line 178)	TinyBase schema — ai_language: { type: "string" }

Recent related changes

• 92172ab41 — "fix: persist language settings properly" (Mar 19, 2026): Fixed language defaults and persistence, added languagePrefixMatch logic so OS-detected defaults aren't unnecessarily written to disk.
• a15c6dce3 → 0c5725a44 — An attempt to translate template section headings to the target language for non-English users was made and then reverted (fix: translate template section headings to target language for non-English users #4072 → Revert "fix: translate template section headings to target language for non-English users" #4120), suggesting localization of generated content is a known area of complexity.
• 36e05f8c2 — Added missing language field for Soniox STT adapter.

Existing infrastructure that could help

1. crates/language/src/detect.rs already has a detect(text: &str) -> Language function using the whichlang crate. It supports 16 languages (including Polish via the broader language family). This could be used to detect the transcript language before enhancement.

2. The EnhanceSystem struct already accepts language: Option<String> — the plumbing from language selection to prompt rendering is fully in place. The main gap is that getLanguage() in enhance-transform.ts only reads the global setting.

3. The TinyBase sessions table does not currently have a language column — one would need to be added to packages/store/src/tinybase.ts to support per-session language storage.

Potential approaches

Option A: Per-session language override (manual selection)

• Add a language column to the sessions table in TinyBase schema
• Add a language picker in the session/note UI (e.g., in the note header area near the enhance button at apps/desktop/src/session/components/note-input/header.tsx)
• Modify getLanguage() in enhance-transform.ts to check session-level language first, falling back to the global ai_language setting

Option B: Automatic language detection

• After transcription completes, run hypr_language::detect() on the transcript text
• Store the detected language on the session or transcript record
• Use it as the default for getLanguage(), still allowing manual override
• Note: whichlang supports 16 languages; Polish (pl) is not directly listed but could be added, or a more comprehensive detection library could be used

Option C: Hybrid (recommended)

• Auto-detect transcript language and pre-fill the per-session language selector
• Allow manual override via the session UI
• Fall back to global ai_language if neither is set

Additional notes

• The same language resolution would need to apply to title generation (title-transform.ts) for consistency.
• The chat system prompt (ChatSystem) also uses a language field (resolved in apps/desktop/src/chat/transport/use-transport.ts), which would benefit from the same per-session awareness.
• The reverted PR fix: translate template section headings to target language for non-English users #4072 (translating section headings) is related context — any implementation should consider how auto-generated template sections interact with the target language.