Improving the quality of RAGLight retrieval can be achieved through 3 dimensions:
- parameterization: Enlargement
kvalues (e.g. set to 8-10) can retrieve more document fragments, but will increase the computation time consuming. It is recommended to start with k=5 for progressive testing - Mode Selection::
- start using
RAT模式pass (a bill or inspection etc)reflectionParameters (recommended values 2-3) add reflection steps to improve logical rigor - adoption
Agentic RAG(used form a nominal expression)max_stepsParameters enable multi-round search optimization
- start using
- embedding model: replace the default vector model with
all-MiniLM-L6-v2High-quality options such asVectorStoreConfigspecified inembedding_modelcap (a poem)provider=Settings.HUGGINGFACE
Note: Model selection should consider hardware performance; large embedded models may significantly increase memory consumption.
This answer comes from the articleRAGLight: Lightweight Retrieval Augmentation Generation Python LibraryThe
May not be reproduced without permission:AI productivity tools " How to optimize the relevance and accuracy of RAGLight search results?
