Week 6 · RAG and evals · about 3 minutes
What you will be able to do
Use hybrid search so Helix catches part numbers with BM25 and paraphrases with vectors.
The idea
BM25 loves exact tokens: SKUs, codes, drug names spelled precisely. Vectors love paraphrases: "blood thinner" ≈ warfarin guidance.
Run both. Merge the rankings. Helix needs both styles of question.
def hybrid(query: str, k: int = 10):
return merge(bm25(query, k=k), vectors(query, k=k))Exact IDs and fuzzy meaning are different retrieval problems — solve both.
Key takeaways
- Keywords catch codes and SKUs.
- Vectors catch paraphrases.
- Merge both lists for Helix.
Part of the Forward Deployed Engineer roadmap. New to the role? Start with what a Forward Deployed Engineer does.