Feat: Translation Memory

[TenzDelek]

Summary

Add Translation Memory (TM) system to optimize translation workflows by caching and reusing previous translations. The system provides both exact matching for instant retrieval and fuzzy matching using PostgreSQL's trigram similarity to surface relevant prior translations, significantly reducing LLM API calls and improving response times for repeated content.

{
  "results": [
    {
      "output_text": "ཁྱེད་རང་ག་དེ་འདུག",
      "from_memory": true,
      "fuzzy_matches": []
    }
  ],
  "metadata": {
    "initialized_at": "2026-04-01T06:25:29.735603+00:00",
    "total_batches": 0,
    "completed_at": "2026-04-01T06:25:29.735603+00:00",
    "total_processing_time": 0
  },
  "errors": []
}

[TenzDelek]

first try to find exact match with the input, if matched ignore the llm step

[TenzDelek]

if fails then we fuzzy match with the input

[TenzDelek]

here the input still go under llm translation

[TenzDelek]

if true, that means exact match successes so we return it without the need to call the llm

[TenzDelek]

fuzzy match logic generated via opus

[TenzDelek]

self reviewed