Thanks! Vector databases store the semantic vector representation of each record and compare it to the query for retrieval, which would give results close to the meaning of the text, but not necessary the text surface. A lexical search, i.e. BM25 and levenshtein distance, seems to work better as translation examples in this case
T-Ragx - Enhancing Translation with RAG-Powered LLMs
Enhancing Translation with RAG-Powered Large Language Models - rayliuca/T-Ragx
cross-posted from: https://lemmy.ca/post/16866615
> Excited to share my T-Ragx project! And here are some additional learnings for me that might be interesting to some:
>
> - vector databases aren't always the best option
> - Elasticsearch or custom retrieval methods might work even better in some cases
> - LoRA is incredibly powerful for in-task applications
> - The pace of the LLM scene is astonishing
> - TowerInstruct and ALMA-R translation LLMs launched while my project was underway
> - Above all, it was so fun!
>
> Please let me know what you think!
T-Ragx - Enhancing Translation with RAG-Powered LLMs
Enhancing Translation with RAG-Powered Large Language Models - rayliuca/T-Ragx
Excited to share my T-Ragx project! And here are some additional learnings for me that might be interesting to some:
- vector databases aren't always the best option
- Elasticsearch or custom retrieval methods might work even better in some cases
- LoRA is incredibly powerful for in-task applications
- The pace of the LLM scene is astonishing
TowerInstructandALMA-Rtranslation LLMs launched while my project was underway
- Above all, it was so fun!
Please let me know what you think!