SENTENCE-BERT · TRANSFORMERS.JS · MAIN /DOCS
Search an ML portfolio by meaning, not only keywords.
A static semantic-search application that embeds portfolio documentation, ranks passages with cosine similarity, and runs entirely in your browser.
LIVE RETRIEVAL DEMO
Document semantic search
Initializing browser model…
Preparing semantic model…
HOW IT WORKS
Static deployment, real Transformer inference
Public documents are processed into section-aware chunks. The app loads precomputed vectors when available; otherwise it generates and caches normalized document embeddings locally. A query embedding is compared with each eligible chunk, and the highest cosine-similarity scores are returned with provenance.
- 01Load
Fetch static corpus and metadata JSON.
- 02Embed
Run all-MiniLM-L6-v2 through Transformers.js.
- 03Rank
Apply filters and cosine-similarity scoring.
- 04Explain
Show source, section, tags, score, and latency.
EVALUATION
Verified evaluation results
Loading evaluation artifacts from data/*.json…
Recall@K
Checks whether a relevant document or chunk appears within the top K results.
Loading… Waiting for recall_at_k_results.jsonMRR
Rewards systems that place the first relevant result near the top of the ranking.
Loading… Waiting for mrr_results.jsonCosine analysis
Inspects score distributions, weak matches, and high-similarity false positives.
Pending review No completed cosine analysis artifact was found.Query latency
Python evaluation reports end-to-end embedding and ranking latency; live searches also report browser latency.
Loading… Waiting for query_latency_results.jsonResponsible use and data safety
This educational demo may return incomplete, outdated, irrelevant, or imperfectly ranked results. Similarity scores are not probabilities. Do not add private company files, internal quality or GCS records, complaint investigations, customer data, emails, credentials, copyrighted material, or personally identifiable information to a public corpus. Review results before using them for decisions.