• English

      • FAQ

      General troubleshooting: Press F12 to open the browser developer tools, switch to the "Network" tab, and check the Response details of the request. Most errors can be identified here.

      What if the translation result is empty, shows only the original text, or returns null?

      Common causes:

      • Configuration Error: The API Key is invalid, or the translation interface parameters are incomplete.
      • Quota/Rate Limited: Account Credits are exhausted, or the request rate is too high and has been temporarily restricted by the interface.
      • Cache Hit Original: Previous original text stored in the cache is being returned directly.
      • Network Restricted: The selected interface (e.g., OpenAI, Gemini, Claude, etc.) is restricted in the current region, or proxy/network anomalies caused the request to fail.

      ✅ Troubleshooting Order:

      1. Verify the API Key and interface settings.
      2. Check account quota, rate limits, and 429 errors.
      3. Disable or clear the translation cache and retry.
      4. Confirm the network environment supports the interface being used.
      5. Inspect the API response in DevTools: press F12 (or Ctrl+Shift+I) → open Network → click "Translate" again → open the latest translation request and check Status (e.g., 401/429/5xx) plus the error details in Response/Preview.

      If only a few sentences failed to translate, you can simply click "Translate" again; the cache will skip completed content and will not deduct fees repeatedly.

      Local model reporting cross-origin (CORS) or connection failure?

      When using local models like Ollama or LM Studio, common causes for failure are browser CORS policies or ad blockers:

      1. Temporarily disable ad/privacy extensions and refresh.
      2. Enable CORS for the local service according to the Translation API Guide (e.g., set OLLAMA_ORIGINS=*, check "Enable CORS" in LM Studio).
      3. If it still fails, check for port conflicts and view the return status code in the Network panel. Company/Campus networks also need to ensure the firewall isn't blocking local ports.

      Running translategemma on Ollama is slow and dropping lines — what's wrong?

      Most likely you've configured translategemma as a regular LLM: picked "Custom (OpenAI-compatible)" and filled in translategemma-4b-it (or similar) as the model name. That path uses the generic LLM pipeline with batching and context markers, and small (under 14B) models struggle to preserve the prompt's structure — you get dropped lines and slower runs.

      ✅ Fix: pick the dedicated "TranslateGemma" service from the dropdown. It uses a purpose-built line-by-line path that matches the Gemma translation model's I/O format and is the most reliable option for local small models. See Translation API Guide → TranslateGemma.

      Why is translation slow?

      The fastest fix: raise "Concurrent Lines" by 50%–100% above the current service's default; lower it immediately on 429 errors. Other tips:

      • Keep the cache on — identical content won't be re-translated (already on by default)
      • Turn on context — AI models only — more coherent, higher throughput
      • Pick a faster service — free GTX runs circles around tiny local models; paid DeepL is faster than Claude Opus
      • For small local models (<14B), turn context OFF: enabling context-aware translation makes tiny models drop lines. Switch to line-by-line concurrent mode — faster AND more reliable

      Defaults vary per service. See Feature Guide → Concurrent Lines.

      Will failed translations retry automatically? What's not retried?

      Yes — network blips, 429 rate-limits, and 5xx server errors retry up to your configured "Retry Count". Failed lines fall back to the original text (never empty), and a red panel at the top of the result lets you retry with one click.

      These errors do NOT retry — retrying can't change the outcome:

      • Bad API key (401 / 403) — fix the key in settings first
      • Timeout — raise "Retry Timeout (seconds)", or pick a faster service
      • Text too long (context length exceeded) — split the file, or lower "Context Lines"
      • Token truncation (max_tokens reached) — split the file, or turn off the maxTokens cap

      Full mechanism: Feature Guide → Failed-Line Retry.

      What if the AI translation quality is unsatisfactory?

      • Restore defaults first: Reset in "Translation Settings" and re-test with a small text segment.
      • Then adjust Temperature:
        • 0~0.3: Strict terminology, requires stability;
        • 0.4~0.7: General scenarios;
        • 0.8~1.0: Allows paraphrasing/more creativity.
      • Enable context-aware translation to improve dialogue coherence.

      The basic translation capability of most online mainstream 70B+ models is generally fine; abnormal output is usually related to temperature or prompt settings.

      Why use a third-party interface to access DeepL?

      DeepL officially prohibits direct API calls from web pages, so a relay channel is needed to forward requests. This channel is only used for forwarding and will not record your data; if you need extreme stability, you can build your own dedicated proxy.

      Is the API Key saved?

      No. The API Key and all settings are stored only in the local browser; no server can access them.

      Why is GTX Web disabled by default?

      GTX Web puts significant pressure on the shared service, so it is disabled by default. If using it locally for personal use, you can manually enable it; enabling a global proxy or unstable networks may cause request anomalies.