4/ For SQL, I use vetted templates, a column whitelist and read-only access. The model is not allowed to generate arbitrary SQL. Keep retrieval and aggregation separate: https://t.co/56FfQAqn8L
1/ A vector search cannot count. If you ask how many suppliers are in the data corpus, it returns the k chunks most relevant to the question. Not a complete set or a reliable count.
3/ I split the paths. Semantic questions go to the vector index. Questions such as "how many", "list all", "sum" etc. go to a small SQL metadata sidecar built during ingest. And a router selects the path.
RRF is a small mechanism, but it removes one fragile parameter from hybrid retrieval. I use it in the open-source Strata-RAG project: https://t.co/56FfQAqn8L
In retrieval-augmented generation, cosine similarity for vector search and full-text BM25 search scores should not be added together. They are complementary, but their raw scores live on different scales.
Reciprocal Rank Fusion avoids raw scores. It sums 1/(k+rank) across both result lists. Items ranked high by both rise. So an unusual score range for one retriever cannot dominate.
The first time I blended two retrievers, hand-weighting the scores was the obvious first idea. A weighted sum can look ok on one dataset, then drift when the corpus changes.
Hey @toptal, looks like there is an issue on video submission during application process. Clicking on 'Complete submission' after successful uploading or recording gives an error