One Data Layer: Relational, Graph and Semantic
02 data layer
Features
02.1One Query Across All Three
A single step filters on fields, matches vocabulary terms, follows a link and ranks by meaning. It returns the records themselves.
- Filter records on their fields.
- Match records by category terms.
- Follow links to related records.
- Rank results by meaning.
- Combine all four in one step.
- Get whole records, not just ids.
02.2Search Declared, Not Built
Mark fields as searchable and the platform splits, embeds and indexes every record, with no vector store to run. A profile can also blend keyword rank with meaning.
- Make any table searchable by meaning.
- Skip running your own vector store.
- Index every record automatically.
- Search long documents in chunks.
- Blend keyword and meaning search.
02.3Scores and the Matching Passage
Each result carries a similarity score that keeps its meaning across tables, plus the passage that matched. A record can also be the query, for “more like this”.
- See how closely each result matches.
- Show users the passage that matched.
- Drop results below a score.
- Compare scores across different tables.
- Find records like this one.
02.4Self-Filing Vocabularies
Category lists match exactly or by meaning, nest child terms under parents, and merge duplicates. Processing files each record under terms and mints a new one when none fits.
- Tag records with shared categories.
- File records under terms automatically.
- Match categories by meaning, not spelling.
- Nest subcategories under parents.
- Merge duplicate categories into one.
- Grow the list when nothing fits.
02.5Swappable Embedding Models
Activating a new embedding model reindexes every table on it. Switching back loses nothing, because the old index is never dropped.
- Switch to a newer embedding model.
- Reindex every table automatically.
- Switch back without losing anything.