- Fixes deleteNamespaceSync deletion of all nested entries.
- Adds hybrid text/vector search (credit:richardsolomou)
- AI SDK v6 support
- Track bandwidth more accurately to avoid going over the limits while deleting chunks.
- Adds /test and /_generated/component.js entrypoints
- Drops commonjs support
- Improves source mapping for generated files
- Changes to a statically generated component API
- Support gateway string IDs for models
- Return usage data from embedding for search/add
- Support text embedding models as a string
- Fix orphaned embeddings when replacing content
- Adds support for AI SDK v5
- Fix orphaned embeddings when replacing content
- ai is a regular dependency
- namespaces can be deleted if there are no entries left in them
- namespaces can be synchronously deleted from an action if there are entries in them
- Allow deleting an entry by key asynchronously or sync
- Deprecated:
.deletefrom mutations is deprecated..deleteis now synchronous for an entry. Use.deleteAsyncfrom mutations instead. - Fix: Delete embeddings when deleting entry
- Fix: Replacing small documents by key no longer leaves them in "pending" state.
- query can be a string or array, instead of separate embedding argument.
- nicer examples & UI to play with
- default chunk sizes are smaller
- EntryFilterValue is now called EntryFilter
- Fixes chunker handling of empty lines
- supports sha-1 content hashing in utility
- default context formatting separates content better for LLMs
- list can take a limit instead of paginationOpts
- findExistingEntryByContentHash is renamed to drop the Existing
- Demote document titles to h2 when auto-generating prompt template
- Rename replacedVersion -> replaced{Entry,Namespace} to match onComplete
- Allow listing documents by status without specifying a namespace (e.g. vacuuming)
- Return replacedAt when listing documents
- Renamed to RAG
- Adds a default chunker, so you can pass
texttoadd[Async] - Adds a
generateTextwith default prompt formatting for one-off generation. - OnComplete handler now has updated status for the replaced & new entry/namespace
- Example showcases prompting as well as searching.
- Add VSearchEntry type for casing vSearchEntry to a type-safe version
- Add SearchEntry type with type-safe access to metadata & filter values
- Allow adding files asynchronously
- Allow passing an onComplete handler to creating entries or namespaces, that is called when they are no longer pending.
- Support generic type-safe metadata to be stored on the entry.
- Updated the example to also show uploading files via http.
- Renamed doc to entry
- Allows passing vectorScoreThreshold to search
- More convenient
textreturned from search - Enables passing in your own embedding parameter to add -> Allows adding (a few chunks) from a mutation.
- Snips console logs
- Vector search over chunked content, with namespaces, search filters, etc.
- You can also gracefully transition between models, embedding lengths, chunking strategies, and versions, with automatically versioned namespaces.
- See the example for injesting pdfs, images, audio, and text!
- List namespaces by status, entries by namespace/status, and chunks by entry
- Find older versions by content hash to restore.
- Add metadata filters for searching.