PotPlayer Ollama Translate

PotPlayer Ollama Translate is a real-time subtitle-translation plugin for PotPlayer, written in Angel Script. It sends each subtitle line to a local or cloud LLM through a custom three-slot prompt template and a rolling bilingual context history, and renders the translation alongside the original so you can watch content in any language without leaving the player. The v3 rewrite builds on v1 with a fully redesigned prompt and context system, a provider-agnostic pipeline, and supports Ollama (native and Cloud), LM Studio REST, and OpenAI- and Anthropic-compatible endpoints.

Role

Solo Developer

Contributions

• Built real-time subtitle translation inside PotPlayer’s extension model — each subtitle line is sent to an LLM, translated on the fly, and rendered next to the original.
• Wrote a multi-provider API abstraction that feeds the same prompt + context through four protocols (Ollama native, LM Studio REST, OpenAI-compatible, Anthropic-compatible), switchable via a single apiFormat field rather than four parallel implementations.
• Designed a rolling bilingual context history: every line is stored as source ⇒ translation and the most recent N entries are re-injected on the next request, so the model uses its own prior output as a style and term anchor instead of translating each line in isolation.
• Built a three-slot prompt-template system (systemPrompt / userPrompt / contextPrompt) with placeholder variables for source language, target language, the current line, and the optional context block — so translation behavior is shaped by editing prompts, not translation logic.

Technical Challenges

• Prompt design in an embedded scripting host. PotPlayer extensions run in Angel Script with no real client-side tooling, so almost all of the translation behavior lives in the prompt — disallowing explanations, forbidding the model from translating the context block, enforcing single-line output, and preserving names/numbers/code are rules baked into systemPrompt rather than enforced by the client.
• Context as a two-axis knob. Real-time translation is a latency-vs-quality dial, not a toggle: contextCount controls how many prior lines are sent per request, contextMaxSize controls how many the buffer keeps; the two are deliberately separate, so a user with a small local model can shrink per-request payload without losing long-window coherence.
• Bilingual history as a style anchor. Storing source ⇒ translation (rather than source only) gives the model its own previous output to mimic, which stabilizes tone, register, and term choice across a scene; the trade is a few extra tokens per request, paid back several times over in consistency.
• Provider isolation so prompts stay single-source. Four API formats expect four request shapes; by keeping request construction as a thin builder and prompt/context assembly above it, every new backend is a builder, not a prompt rewrite.

Lessons Learned

• The v1→v3 rewrite’s real leverage was the prompt and context system, not the API plumbing — the moment translation behavior moved into editable templates, the plugin stopped being “Ollama with extras” and became a general LLM translator.
• Bilingual rolling context was a small change with an outsized effect; exposing contextCount and contextPrompt to the user (not just temperature) turned the plugin from “configure the model” into “configure the translation”.
• Building provider support as a single apiFormat switch from the start — instead of an Ollama-only plugin later retrofitted for “other APIs” — made every additional backend a config change rather than a rewrite.

Source: GitHub · Releases: latest