Glossary, Streaming, and Line Breaking

Use this page when a long manga series must keep names, places, and skills consistent, when you want to see incremental output while a translation is in flight, or when translations should wrap according to the original line count. It documents automatic glossary extraction (translator.extract_glossary), streaming (translator.enable_streaming), and AI line breaking (render.disable_auto_wrap). Extracted terms are written back to the custom prompt file; streaming changes request transport only; AI line breaking uses a line-break prompt plus original_region_count so the model returns [BR] markers.

It does not cover translator or target-language selection (see Translator selection), the full history and prompt-composition picture (see Context and prompts), or renderer-side auto wrapping, semantic breaking, and punctuation trimming (see Typesetting and rendering).

When to use it

• There is no translator.glossary configuration key in the current config model; the glossary lives under a glossary key inside the custom HQ prompt file (translator.high_quality_prompt_path) and is written back by the auto-extraction feature.
• translator.extract_glossary is the automatic glossary-extraction switch; it enters the extraction branch only when the custom HQ prompt loads successfully, otherwise the switch alone still uses normal translation.
• translator.enable_streaming (the translator.stream shorthand in the task brief) changes request transport only for OpenAI/Gemini translators, including HQ modes; it does not change prompts, context, glossary extraction, or the final translation.
• render.disable_auto_wrap is displayed as “AI Line Breaking”. It drives both the translator-side line-break prompt and the renderer-side [BR] forced-wrap semantics; renderer auto wrapping itself is covered by the typesetting page.
• OPENAI_GLOSSARY_PATH (displayed as “Glossary Path”) is a legacy environment-variable-backed glossary path and is separate from the location where extract_glossary writes terms (the custom prompt file).

Set it in the desktop app

Enable glossary extraction and streaming in Settings

1. Open “Settings” and select the “Translation” group.
2. Toggle “Auto Extract Glossary”. When enabled, the description panel shows “Auto-extract proper nouns (names, places) from translations to ensure consistency in long manga series.”
3. Toggle “Enable Streaming”. The description panel explains the difference between streaming and non-streaming requests.
4. Glossary extraction requires a parseable prompt file selected under “Custom Prompt”; without one, the switch does not produce the extraction branch.
5. Open “Settings” → “Typesetting” and toggle “AI Line Breaking” to enable the AI line-break prompt.

Inspect the glossary in Prompt Preview

Open “Prompt Management”, select a custom prompt file, and click “Prompt Preview”. If the file contains a glossary key, the preview shows a “Glossary” section with a total entry count and per-category tabs for Person / Location / Org / Item / Skill / Creature; when empty it shows “No glossary entries”.

Parameters and options

For how each parameter's UI name, storage key, and default value map to each other, see UI Options Reference.

Auto Extract Glossary

• Control: toggle.
• Location: Settings → Translation.
• Options: on or off.
• Default: false.
• Mechanism: when enabled, every attempt appends glossary-extraction rules and an extended output format that requires new_terms. Proper nouns such as names and places are extracted from the response and written back under the glossary key of the custom prompt file. A new original creates a same-name first aliases form and stores its translation in that alias's translations; an existing original accepts only genuinely new aliases when overwrite: true, while false or an omitted field leaves it unchanged. An alias that already exists is discarded as a whole, so AI never adds a second translation to it. AI deltas use the original/category/aliases shape; conditions, descriptions, overwrite flags, and other unknown fields are ignored on write-back. The next batch reloads the file, forming the “extract → write back → carried by the next request” feedback loop. The extraction branch runs only when a custom prompt loads successfully; otherwise the switch alone still uses normal translation. Write-back modifies the user prompt file, so sanitize it before sharing.

flowchart LR
    subgraph Off["extract_glossary off or custom prompt invalid"]
        N1["Base system prompt + standard translations output format"] --> N2["OpenAI/Gemini request"]
        N2 --> N3["Parse translations only, no write-back"]
    end
    subgraph On["extract_glossary on and custom prompt valid"]
        G1["Base system prompt + glossary rules + new_terms output format"] --> G2["OpenAI/Gemini request"]
        G2 --> G3["parse_hq_response → translations + new_terms"]
        G3 --> G4["Deduplicate by category and write back to the glossary key"]
        G4 -. "Next batch reloads the file; glossary is flattened into the custom prompt" .-> G1
    end

The switch only changes prompt content and write-back behavior; translation-count validation, retries, and candidate rotation still run as usual.

Enable Streaming

• Control: toggle.
• Location: Settings → Translation.
• Options: on or off.
• Default: false.
• Mechanism: when enabled, OpenAI/Gemini translators, including HQ modes, prefer the unified streaming transport to receive incremental responses in real time, with incremental previews in the console/logs; when disabled, they always use standard non-streaming requests. If a streaming request raises an exception (for example an endpoint that does not support streaming), this attempt automatically falls back to a normal non-streaming request without interrupting the task.

flowchart LR
    R["Translation batch request"] --> S{"enable_streaming?"}
    S -->|off| N["Standard non-streaming request, full response at once"]
    S -->|on| T["Unified streaming transport (incremental/accumulative/repeated chunk normalization)"]
    T --> P["Incremental JSON preview in console/logs"]
    T --> E{"Streaming raised an exception?"}
    E -->|yes| F["Fall back to a normal non-streaming request for this attempt"]
    E -->|no| D["Aggregate full text + finish_reason"]
    N --> V["Validate and parse translations"]
    F --> V
    D --> V

The fallback affects a single attempt only; if the endpoint keeps failing, retries and candidate rotation take over as usual.

AI Line Breaking

• Control: toggle.
• Location: Settings → Typesetting.
• Options: on or off.
• Default: false.
• Mechanism: when enabled, the translation side adds the line-break prompt to the system-prompt prefix and asks the model to output [BR] markers according to the original line count; the renderer treats [BR] as a forced wrap during typesetting. It interacts with options such as “AI Line Break Check”; the full behavior is in the typesetting-and-rendering page. Replace-translation mode forces AI line breaking and the strict layout. For single-line regions (N=1), if the model still returns [BR]/<br>/【BR】, the markers are automatically cleaned into a single line during line-break validation; this cleanup does not depend on the “AI Line Break Check” switch.

flowchart LR
    subgraph Off["disable_auto_wrap off"]
        O1["No line-break prompt loaded"] --> O2["User prompt without original_region_count"]
        O2 --> O3["Rendering: auto-wrap layout"]
    end
    subgraph On["disable_auto_wrap on"]
        A1["Load system_prompt_line_break.yaml"] --> A2["Line-break prompt enters the system-prompt prefix"]
        A1 --> A3["Attach original_region_count per region"]
        A2 --> A4["OpenAI/Gemini output [BR] markers"]
        A3 --> A4
        A4 --> A5["Rendering: forced wrap on [BR]"]
        A5 --> A6{"check_br_and_retry and region count ≥ 2?"}
        A6 -->|translation lacks [BR]| A7["Trigger retry"]
        A6 -->|ok| A8["Proceed to next stage"]
    end

The full renderer branches for auto wrapping, HanLP semantic breaking, and punctuation trimming are covered by the typesetting page; this guide only explains how the line-break prompt enters the translation request.

How translation requests are handled

Glossary extraction, merge, and feedback

Glossary extraction relies on two facts: the extraction branch runs only when a valid custom prompt is loaded and extract_glossary is on, and new terms are written back to the custom prompt file (high_quality_prompt_path), not to the file pointed to by the OPENAI_GLOSSARY_PATH environment variable. The written glossary key is organized by standard categories. Existing originals stay unchanged by default; only entries that explicitly allow automatic additions receive genuinely new aliases. An alias that already exists is discarded, even when the AI proposes a different translation. Automatic merging does not modify authored translations, conditions, descriptions, overwrite flags, or other authored content. The preview page shows these entries in per-category tabs. The feedback loop takes effect only after the next batch reloads the file; it never mutates an already-built request.

Unified streaming transport and fallback

The stream preview affects console/log output only; the aggregated full text still goes through the shared response validation and parsing. A streaming failure does not switch translators or candidates; it only falls back to a normal request for that attempt.

AI line-break prompt and [BR] markers

The user prompt carries original_region_count, which the renderer uses to judge whether the [BR] count in the translation matches the original line count; check_br_and_retry retries only translations with ≥2 regions that are missing [BR]. For single-line regions (original_region_count=1), even if the model returns [BR]/<br>/【BR】, the markers are automatically cleaned into a single line; this cleanup depends only on disable_auto_wrap, not on the check_br_and_retry switch.

Models, network, and quality

• extract_glossary is tightly coupled to high_quality_prompt_path: without a valid custom prompt there is no extraction branch and no term write-back.
• enable_streaming is independent of prompts, context, and glossary extraction; it changes transport only.
• disable_auto_wrap affects both translation and rendering stages; the full combination of optimize_line_breaks, semantic_linebreak, remove_linebreak_punctuation, and check_br_and_retry is on the typesetting page.
• Streaming, RPM, ordinary retries, and API candidate rotation stack on the same request path; they do not change glossary or line-break content.
• Glossary and prompt files may contain business text. Remove request bodies, glossary entries, paths, and credentials before sharing logs, request exports, or debug directories.