Neural vs keyword search

Vulgate offers two complementary search modes. Picking the right one for the question at hand makes a big difference to result quality.

Neural search

Neural search converts your query and every passage in your Library into a vector — a numeric fingerprint of meaning — then returns the passages whose meaning is closest to your query. It looks for what you mean, not what you said.

Use neural search for

• Conceptual questions — "the early Church's view on civic authority", "how the council described liturgical music".
• Open-ended exploration — when you're not sure what terms an author used.

Strengths

• Tolerates synonyms, paraphrases, and translation.
• Returns the most semantically relevant passages first.
• Works even when your query and the source use no shared vocabulary.

Limits

• May surface results that sound close but aren't exactly what you wanted. Use the snippet and the Preview pane to verify.
• Not the right tool for proper nouns or exact quotations.

Keyword search

Keyword search returns passages that literally contain the words and phrases you typed. It's fast, deterministic, and easy to reason about.

Use keyword search for

• Proper nouns and named entities — "Athanasius", "Constantine", "Trent".
• Exact quotations — when you remember a turn of phrase and want to track it down.
• Technical terms — Greek transliterations, Latin word forms, codified acronyms.
• Verbatim verification — confirming that a specific word is in a document.

Strengths

• Predictable: if the words are there, they show up.
• Fast.
• Easy to scan results by snippet.

Limits

• Misses synonyms and paraphrases.
• Misses translations: a search for "council" will not find a Latin passage that uses "concilium" unless you search for the Latin term directly.

Combine both

It's normal to alternate. Many researchers start with a neural query to find the relevant area, spot a key term in the results, then switch to keyword search to track every occurrence of that term across the Library.

Behind the scenes

Vulgate's neural search runs over embeddings generated during ingestion. Keyword search runs over a full-text index built at the same time. Both indices stay in sync automatically when documents are edited or re-ingested.

If you want a synthesized answer on top of these retrieval engines, switch to Chat with your Library — the AI retrieves from your Library and writes an answer with citations.