Semi-autonomous translation workflow with Agentic RAG
The article A semi-autonomoous translation workflow with Agentic RAG explains how an AI agent can manage an entire professional translation workflow using Agentic Retrieval Augmented Generation (RAG). The example agent, Dani, is a semi-autonomous system that not only translates text but also retrieves domain-specific information, applies client-specific terminology and tone, performs quality checks, and adapts to unexpected situations.
Because large language models lack up-to-date, domain-specific, or client-aligned knowledge, RAG is used to retrieve relevant information—such as terminology stored in a knowledge graph, tone-of-voice guidelines, and forbidden or preferred terms—at generation time. Unlike classic RAG pipelines, Agentic RAG allows the agent to decide which tools to use and in what order, making the workflow more flexible and intelligent.
This approach is especially useful for translation tasks where terminology changes frequently, quality requirements vary, and precision is critical (e.g. medtech). By combining an LLM with retrieval systems, domain-specific tools, and memory, Agentic RAG enables AI agents like Dani to deliver high-quality, consistent, and context-aware translations with relatively little implementation effort, using frameworks such as Smolagents, LlamaIndex, or LangGraph.
Coreon’s Multilingual Knowledge Graph offers a tool that is more than a terminology store—by offering term annotation and domain classification, the agent can extract more precise information about terms and generate an even more accurate translation.
ISO TBX going RDF
TBX is the known standard for sharing terminological resources and is supported by many – including Coreon. Although XML has many benefits as an exchange format, it is not the usual format used in the semantic web and knowledge graph communities.
Instead, RDF is the established syntax for representing and sharing knowledge and vocabularies to be used in Semantic Web and related contexts. An RDF representation of TBX also makes terminology resources easily available in these contexts, which facilitates the work for the developers involved. Technically speaking: It makes terminology resources available as an RDF graph. This means that terminology can be incorporated into an RDF-based ecosystem immediately when deploying knowledge and language resources.
Coreon has always advocated the use of methods and technologies to make terminology resources available for use in contexts beyond documentation and translation. RDF helps overcome barriers. The first milestone was achieved in 2021 with the ELG MKS2LLOD project, in which we demonstrated the process of mapping concept-oriented terminology resources into RDF triples.
We contributed these experiences to the ISO TBX Working Group, which has now been turned into ISO Technical Recommendation 24633-3. It is currently in draft status (“ISO 20.00”) heading towards approval. The draft RDF vocabulary is already available at the ISO resources page. In the name of the ISO Working Group Michael Wetzel, Coreon MD and Head of Product, welcomes any comments on the draft vocabulary.
Events: tekom/tcworld, Language Intelligence, GenAI in Localization
We are looking back at three successful events this winter …
- Carina, Coreon’s data scientist, shares her impressions during the tekom conference and the emphasis on AI, LLMs …
- Jochen, our CEO, summarises upon the LI opening panel that the industry must re-invent itself, must evolve from processing content to mastering knowledge.