Skip to content

Iterative multi-hop retrieval-augmented QA that updates the query embedding algebraically and stops when retrieval utility is exhausted

License

Notifications You must be signed in to change notification settings

mauriziotirabassi/isce_iterative_rag

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

18 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Iterative Semantic w/ Cross Embedding (ISCE) RAG

This repository contains a research notebook exploring a novel method for multi-hop question answering.
The method works entirely in embedding space: after each retrieval step, we algebraically update the query vector to remove already-covered information. We stop retrieving once the information gain plateaus, then pass the collected passages to a generative model to produce the final answer.

Key Idea

The core principle is that the question already contains information about the necessary hops to be performed.

  • Embed the question with a sentence embedding model.
  • Retrieve passages from a dense index (hnswlib).
  • Update the query embedding after each hop:
    • Subtract or adjust with the passage embedding to represent “remaining information.”
  • Check if retrieval scores have plateaued (“semantic exhaustion”).
  • Stop when exhausted, and generate the final answer from the gathered passages.

This avoids the need for a learned controller and makes multi-hop retrieval lightweight and unsupervised. Instead of relying on supervision for when to “hop,” ISCE shows that simple algebra in embedding space can control multi-hop retrieval effectively.

Repository Contents

  • isce.ipynb — the main notebook with method, experiments, and results.
  • Dataset used: HotpotQA (subset).
  • Images of the blueprint and logic in the notebook.

Results (HotpotQA, dev subset)

  • Baseline RAG: F1 ~0.45, EM ~0.38
  • ISCE variants: F1 up to ~0.48, EM up to ~0.41

About

Iterative multi-hop retrieval-augmented QA that updates the query embedding algebraically and stops when retrieval utility is exhausted

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •