MISSING BRIDGE FILLED
From embeddings to vector stores to retrieval-ready results.
The repo already taught embeddings and later built a retrieval system. What it did not show cleanly enough was the middle: how vectors become an index, how metadata changes the candidate pool, and how ranked results become citation-ready payloads.
EMBEDDING MAP
Drag the query through the space.
The circles are chunk embeddings. The gold point is the query. The results panel shows raw semantic ranking beside the filtered ranking.
Drag the gold query point
Dimmed points are excluded by the active filter
Semantic only
Nearest neighbors before filtering
Filtered
Nearest neighbors after metadata
THE BRIDGE
Three things the old sequence left implicit.
These are the concepts that learners need in hand before Chapter 10 becomes straightforward.
Embeddings are not retrieval
An embedding is just a vector until it stays attached to a chunk, an ID, and metadata.
Vector stores are contracts
The backend matters less than the contract: insert vectors, search by similarity, preserve metadata, return ranked results.
Ranking is not enough
To become retrieval, the result still needs a snippet, a parent document, and enough structure to support filtering and later citation.
NOTEBOOK PATH
Use the bridge notebooks before Chapter 10.
The new sequence is simple: embeddings, vector stores, retrieval bridge, ranking lab, then `10.0` on evaluation and reranking before the wider capstone.