CorpusLens: finding a passage when you remember the idea

Sometimes I remember what a passage says but not its wording. A word search can miss it, especially across translations. Searching one verse at a time can miss it too: an argument or story may take several verses to make its point. CorpusLens is my English and Spanish search project for those cases.

Index the text at several sizes

The first source is the New World Translation EPUB in each language. An adapter extracts books, chapters, verses, and section outlines into a versioned snapshot. Before indexing, a coverage audit checks that the text spans are represented in search chunks, with no character gaps. The recorded snapshot covers 66 books and 1,189 chapters in each language; verses absent or present only as notes in the source are recorded explicitly rather than filled in.

Each chapter becomes four search layers: individual verses, overlapping windows of three verses, overlapping windows of five verses, and sections from the source outline. A section’s heading and location are added as embedding context, while the verse text remains separate. The verse layer helps locate a precise match; the larger layers carry the context needed for a longer thought.

Bring different kinds of matches together

An exact citation such as John 3:16 takes a direct route to the referenced verse. For other queries, the search service can retrieve both lexical matches and vector matches. Its weighted reciprocal-rank fusion combines their positions in the two result lists, rather than comparing their raw scores. Overlapping chunks are then grouped into passage results so a verse found in several layers does not appear as several unrelated answers.

The vector pipeline uses 1,024-dimensional text-embedding-3-large embeddings and Oracle’s vector storage; Oracle Text supplies the lexical side. The service can continue with lexical results when an embedding is unavailable. That fallback matters because a useful search result should not depend on every external step succeeding.

Try the English app or Spanish app. The implementation is on GitHub.