Build an Embedding-Based Semantic Search for a Legal-Document Corpus
Overview
What this challenge is about.
Embed the 380k-document corpus using a multilingual sentence-transformer (e.g. multilingual MPNet or LaBSE). Store embeddings in FAISS or pgvector. Build a search service that returns top-K by cosine similarity. Evaluate on a held-out query set (around 200 queries with labeled relevant cases). Metrics: recall-at-10, mean-reciprocal-rank, P@5, query latency. Compare against the existing BM25/keyword baseline. Propose a hybrid (BM25 + semantic reranker) production architecture with cost/latency trade-offs. Deliver: indexing + search code, 6-page evaluation report, hybrid-architecture spec (4 pages), example queries demoing failure modes of each approach.
The Brief
What you'll do, and what you'll demonstrate.
Build a semantic-search service that beats the keyword baseline on recall-at-10 and ships as a hybrid architecture.
Earning criteria — what you'll demonstrate
- Use pre-trained sentence-transformers for cross-lingual semantic search
- Compare semantic vs keyword retrieval with proper IR metrics
- Design hybrid retrieval architectures with reranking
- Translate evaluation results into production architecture decisions
Program Fit
Where this fits in your program.
Sharpens the same skills your degree expects you to demonstrate.
Skills
Skills you'll demonstrate.
Each one shows up on your verified credential.
Careers
Roles this prepares you for.
Real titles. Real skill bridges. Pick the one closest to your trajectory.
Career mappings coming soon.