All projectsAvailable for DFMA, industrial design, and mechanical implementation
JR PEREZ
AI research
First author, ICML 2025 Workshop on Vector Databases

Entropy-Weighted Retrieval

Search that decides, query by query, whether to trust keywords or embeddings more.

Aim

Improve the retrieval step of AI question answering by weighting keyword search and embedding search for each query instead of using one fixed blend.

Tools
  • BM25 keyword search
  • FAISS vector search with MiniLM embeddings
  • Llama 3
  • LLM-as-a-judge evaluation
Skills
  • AI analysis
  • Machine learning
  • Information retrieval
  • Statistical testing
  • Writing and publishing a paper
Outcome

A first-author paper at the ICML 2025 Workshop on Vector Databases. Per-query weighting beat a fixed blend on TriviaQA and held level on HotPotQA.

Synthetic scores illustrate the method. The last frame is the paper's Figure 1.
Idea

A retriever whose scores are sharply peaked is more confident. Its entropy sets how much weight it earns for that query.

Update rule

ws = (1 − Ĥs) / ((1 − Ĥs) + (1 − Ĥd)), iterated until |Δws| ≤ ε.

Next project, 08 of 11 Optical Limiting With Colloids DoD funded physics
Available for DFMA, industrial design, and mechanical implementation.
Inquirejr@smartflowinventory.com