Context-enriched LLMs
Similar translation units retrieved from your TM let small open-weight models rival a frontier model: with three examples, Gemma 3 4B beats Gemini 3 Flash's unaugmented ChrF++.
A machine translation research prototype
Context-augmented machine translation for CAT tools
Echo bridges machine translation (MT) and computer-assisted translation (CAT) using smaller, context-enriched LLMs. Leverage translation memories, termbases, and recent edits to improve quality while keeping data private.
Highlights
Self-updating memory
Translator corrections flow back into the system automatically—no manual TM uploads required.
CAT tool integration
Trados TranslationProvider plugin retrieves context and suggestions directly in the translator’s workflow.
Privacy-first
Local-first and on-prem friendly. Keep sensitive data in-house while improving quality.
How it works
Ingest
Upload TMX/SDLTM and termbases. Content is embedded into Qdrant for retrieval.
Translate
REST API + Trados plugin call LLMs (OpenAI, Anthropic, etc.) with relevant context.
Improve
Edits are captured and cached (FIFO per session) to increase consistency and quality.
Roadmap
Prototype
- DoneStage 1: End-to-end demo (API + plugin) and basic metrics
- DoneStage 2: Automatic feedback to memory
- DoneStage 3: Projects & auth basics
- In progressStage 4: Session cache (needs online backend)
- PlannedStage 5: HyDE + new models + evaluations
Research
Published at UNLP 2026: retrieved examples improved ChrF++ and COMET for every model tested, from Gemma 3 4B to Gemini 3 Flash. Next: more domains and a random-retrieval control.
Metrics
From Improving Domain-Specific Translation from English into Ukrainian with Retrieval-Augmented Generation (UNLP 2026): 258 test sentences from a legal and military EN → UK corpus, each translated with k similar pairs retrieved from 2,323 indexed ones. Every gain over k = 0 is significant at p < 0.01 (paired bootstrap, Holm-corrected).
Gemma 3 4B, 3 examples+0.076COMET
Gemma 3 4B, 3 examples+21.39ChrF++
Gemma 3 27B vs Gemini 3 Flash, both at k = 579.14vs 77.64 ChrF++
Augmented runs significant16 / 16p < 0.01
| ChrF++ | k = 0 | 3 | 5 | 10 | 25 |
|---|---|---|---|---|---|
| gemini-3-flash | 62.47 | 77.49 | 77.64 | 78.62 | 78.76 |
| gemma3-27b | 60.73 | 77.65 | 79.14 | 79.82 | 79.61 |
| gemma3-12b | 54.85 | 71.19 | 73.34 | 74.94 | 75.23 |
| gemma3-4b | 46.87 | 68.26 | 68.28 | 68.81 | 59.09 |
| COMET | k = 0 | 3 | 5 | 10 | 25 |
|---|---|---|---|---|---|
| gemini-3-flash | 0.896 | 0.931 | 0.934 | 0.935 | 0.936 |
| gemma3-27b | 0.875 | 0.920 | 0.928 | 0.927 | 0.931 |
| gemma3-12b | 0.864 | 0.906 | 0.914 | 0.911 | 0.916 |
| gemma3-4b | 0.812 | 0.888 | 0.892 | 0.892 | 0.885 |
One narrow domain: the paper makes no claim that these gains carry over to unrelated documents. Code, prompt and de-identified outputs.
Who is this for?
Translators
Faster, more consistent output within Trados; no new tools to learn.
Agencies
Quality and throughput improvements with existing TMs and workflows.
Enterprise & government
On-prem options for sensitive content; clear path to compliance.
Academia
Results published at UNLP 2026, with the code, prompt and de-identified outputs released for reproduction.