Feature Request: Adding Topic Tagging Stages to RAG Pipeline for Enhanced Vector Similarity Search

[monks1975]

Problem Statement
Currently, the RAG pipeline excels at generating contextually relevant responses. However, there's room for improvement in the vector similarity search phase. It could benefit from additional contextual signals to home in on the most relevant document chunks.

Proposed Solution
I suggest incorporating tagging stages at two key points in the RAG pipeline:

Chunk Tagging: After the document retrieval stage, each chunk could be tagged with topic tags. This can be done by feeding each chunk through another language model stage specialised in topic classification.

Query Tagging: Just before prediction, the input query can also be tagged with relevant topics, perhaps using the same specialised language model as above.

These tags can then be utilised to focus the vector similarity search, offering a more nuanced and contextually rich result set.

Benefits
More focused document retrieval, reducing noise.