Retail Search

v0-cranfield-opensearch-baseline

Live Cranfield baseline

Phase 2 - Search Explain (Cross-Domain)

How Phase 2 retrieves. Tier-1 datasets rank with live BM25; the ARCH-0.5 core adds the BGE hybrid, embedding the query at request time with Cloudflare Workers AI, then blending dense kNN with lexical scores. Run a query to see the live stages.

All phases

Explain a live query (ARCH-0.5 BGE hybrid)

Runs the live arch-0.3-bge pipeline and shows each retrieval stage for your query.

Run explain to see the live BGE hybrid retrieval flow.

Embedding space

Your query and its nearest documents are 768-dimensional BGE vectors. Here they are projected to 2D (PCA) - the query is the star, documents are dots sized and shaded by how close their meaning is to the query. Closer = more semantically similar.

Run explain above to plot the query and its neighbours in embedding space.

Retrieval flow reference

ARCH-0.5 core (BGE hybrid). Tier-1 BEIR datasets use stages 3-4 only (lexical BM25).

  1. 1Query embeddingWorkers AI @cf/baai/bge-base-en-v1.5, retrieval prefix applied (deterministic encoder, no generative LLM)
  2. 2Vector kNN768-dim cosine kNN over the BGE index (k=50)
  3. 3LexicalBM25 multi_match over title + text/abstract
  4. 4Hybrid blendmin-max linear blend of dense + lexical, then top-k