andrewyng/translation-agent
Translation Agent
Translation Agent is a Python-based system that uses a large language model to translate text through a multi-step agentic workflow. Rather than producing a single output, it generates an initial translation, then prompts the same LLM to critique its own work and produce improvement suggestions, and finally refines the translation based on that self-critique. This reflection-driven iterative refinement loop is the core mechanism for improving translation quality without requiring human feedback or additional training data.
The system distinguishes itself through two key capabilities. First, it enforces consistent terminology by injecting a user-defined glossary directly into the LLM prompt at every stage of the workflow, ensuring that names, technical jargon, and other specified terms are translated uniformly across the entire document. Second, it allows users to steer the translation style by adjusting prompt parameters that control formality, handle idioms, and target specific regional dialects, giving fine-grained control over the output's tone and linguistic register.
The entire workflow is orchestrated by a lightweight Python script that sequences stateless LLM calls and manages prompt templates, with all context passed explicitly in each prompt rather than maintained in memory. This stateless, prompt-based architecture keeps the system simple and portable, requiring no dedicated runtime or framework beyond the Python script and access to an LLM API.
Features
- Reflective Translation Agents - Provides an agentic workflow that translates text by generating, reflecting on, and refining translations using an LLM.
- LLM Translation Workflows - Decomposes translation into discrete stages of generation, reflection, and refinement via separate LLM calls.
- Iterative Translation Refinement - Implements a sequential loop of translation, self-critique, and polishing to increase linguistic precision.
- Self-Critiquing Pipelines - Implements a multi-step pipeline where the LLM critiques its own output to produce higher-quality translations.
- Translation Style Parameters - Controls formality, dialect, and idiom handling by adjusting system prompt parameters for translation calls.
- Translation Self-Critique - Uses the LLM to critique its own translation output and generate improvement suggestions before finalizing.
- Self-Critiquing Pipelines - Implements a multi-step pipeline where the LLM critiques its own output to produce higher-quality translations.
- Style Parameter Prompting - Steers output formality, handles idioms, and targets regional dialects through customizable prompt parameters.
- Regional Dialect Overrides - Merges terminology bundles to override translations for regional dialects and adjust formality.
- Reflective Text Translation - Translates text by generating an initial version, reflecting on it for improvements, and refining the output.
- Translation Glossary Enforcement - Enforces consistent translations for names and technical jargon by including a glossary in the prompt.
- Translation Style Steering - Adjusts formality, handles idioms, and targets regional dialects through customizable translation prompts.
- Translation Critique Pipelines - Sequences discrete stateless LLM calls for translation, critique, and refinement in a pipeline.
- Translation Style Parameters - Provides prompt parameters to control formality, dialect, and idiom handling in translations.
- Translation Term Mapping - Overrides automated translation outputs using a glossary to ensure consistent terminology across documents.
- Prompt-Injected Glossaries - Enforces consistent terminology by injecting a user-defined glossary directly into the LLM prompt at every workflow stage.
- Prompt-Injected - Injects glossary terms into prompts to enforce consistent translation of names and jargon across outputs.
- Reflective - Demonstrates a reflection-based workflow for improving machine translation quality through iterative refinement.
- Reflective Translation Workflows - Implements a reflective workflow where the LLM generates, critiques, and refines its own translations in a multi-step process.
- Style Parameter Prompting - Provides adjustable prompt parameters to control formality, handle idioms, and target regional dialects in translations.
- Python Script Orchestrators - Ships a lightweight Python script that sequences stateless LLM calls and manages prompt templates for translation workflows.
- Python Script Orchestrators - Ships a lightweight Python script that sequences stateless LLM calls and manages prompt templates for translation workflows.
- Translation Workflows - An agentic workflow for nuanced, multi-parameter translation tasks.
