Nomic Embed Image: Multimodal Embeddings in Atlas
Atlas supports mapping image datasets using the Nomic Embed Vision model. Datasets with images can be stored in Atlas using the map_data function from the Nomic Python SDK.
Topics for a highlighted datapoint are now visible in the data sidebar
Fixed bug in saving topic labels which get updated during session
Improved reliability of web upload for JSON and JSONL files
Improvements to lasso selections in non-embedding data spaces
Revamp of the map UI for better use of space and ease of use
Topic labels can now be edited by map admins from within the map page; changes to topic labels are immediately reflected in the map and propagated to the server
The ability to update an Atlas index is temporarily deprecated due to critical flaws in it's functionality and data consistency guarantees. Please message support@nomic.ai for details on when the feature will become available again.
Range slider improvements, map projection improvements
Range Slider now supports zooming in and out for more precision, better UI for settings dates and minor data display updates
Improved 2D layout — nomic-project-v1 will now auto-infer hyperparameters based on your dataset by default, resulting in layouts that better capture local and global embedding structure.
Datasets embedded by Nomic now show the embedding model used to create them
Various bug fixes, including more stability on the data upload page
Nomic Embed: The flagship Nomic Atlas embedding model is now launched as an inference endpoint and as the default embedding model in Nomic Atlas.
Topic Label Generation Improvements: The topics labels assigned to detected topics in your datasets have been improved for quality, consistency and relevance. Remember, you can access topics and labels programatically with the Nomic Python Client.
Embedding Inference Optimizations: Lowered P75 latency of embedding inference to 200ms at batch size 100 and seq length 2048.
Browser Data Upload Improvements, access to points, new URLs
Browser Data Upload: Columns are now editable, patched several case bugs involving duplicate names, etc.
It's now easier to isolate a single point on the map--just click and you'll get
immediate access to the most important information about your point,
with the full information available on the left. There's also a handy way to
add it to the current selection--watch here for more operations on
individual items in your dataset!
Atlas uses slugs instead of UUIDs even for maps--the default place to view your dataset
map is now at https://atlas.nomic.ai/my_org/my_project/map. If you've built multiple maps,
they're still accessed by UUID: but most users share just the most recent map they have,
and this makes it easier to understand your links and work with them from the python client.
The in-browser project creation interface is now more reliable and
returns better information about errors in uploaded CSVs.
Composable Selection: All Nomic Atlas data selections are now composable at the ten's of million points scale. Build highly expressive selections over your datasets in the web browser with metadata filtering, regex search and semantic region lasso'ing. Try searching your Atlas dataset, and combine it with a lasso over some of the results.
Atlas now allows you to combine multiple searches together at the same time--lassos, regular expressions searches, and/or
categorical filters. Choose whether to use "any" or "all" to combine multiple filters together.