Routing does not generate a database filter
The notebook maps catalog choices to predeclared filter expressions. Do not replace that mapping with arbitrary model-produced query text.
Inspect the supporting source ↗REPOSITORY ANALYSIS
bootcamp
Runnable search tutorials using Gemini embeddings, Milvus retrieval, and Jev judgments for reranking, filtering, and routing.
E2 · source reviewedPublic material or source evidence was reviewed. A fixed-version source review requires checking a pinned revision; this is not a reproduction, performance endorsement, or safety guarantee.
- Stars
- 2,445
- Forks
- 683
- License
- Apache-2.0
- Last commit
- 2026-09-24
What Jev does here
Jev judges declared query routes and candidate evidence, while notebook code maps fixed choices to filters and validates all requested answer IDs.
Shows retrieval, semantic reranking and deterministic filter policy as separate stages on a small synthetic dataset.
Fixed-version documentation and implementation material were inspected on 2026-10-24. No notebook cells or API credits were used. The synthetic document set is not evidence of production retrieval quality.
What task does this project fit?
Learn where typed judgments fit around a vector-search pipeline.
Material analysed：
From input to action
- Gemini embeds synthetic documents and queries; Milvus performs retrieval.
- Jev classifies query intent or checks candidate evidence with declared questions.
- Code maps fixed choices to filters and applies local thresholds or fallback.
Implementation details that matter
Embedding consistency is a separate requirement
The notebook keeps model, dimensionality and input formatting consistent and embeds documents separately. A Jev reranker cannot repair vectors built under incompatible settings.
Inspect the supporting source ↗When to choose it
Choose it for a small educational retrieval stack. Production adoption needs real relevance labels, explicit privacy decisions and migration rules when the embedding setup changes.
How to inspect it yourself
Read the notebook as JSON or in a notebook viewer without executing cells. Follow judge response-key validation, then the catalog-to-filter mapping; confirm documented thresholds are starting values rather than calibrated defaults.
Limits of this review
- No notebook cells or API credits were used.
- The synthetic document set is not evidence of production retrieval quality.