Categories
Self-Hosted Machine Learning Data Visualization Research Tools Vector Search Data Exploration Embeddings Text ClusteringStars
Forks
Watchers
Developer links
Latent Scope
Survey answers, support tickets, and documents become an interactive map where similar rows sit together, so patterns show up before anyone reads a single row. Latent Scope imports a CSV, Parquet, JSON, JSONL, or XLSX file in the browser, then walks through five setup steps: embed the text with a local or hosted model, project the vectors to two dimensions, cluster the points, label every cluster, and publish the result as a scope. The Explore view pairs a searchable table with the map, so you can type a phrase to run a nearest-neighbor search, browse cluster lists, and open any row's full text. Setup keeps every experiment in a gallery with thumbnails and quality scores, which makes comparing parameters painless. Image collections get a zoomable image map, and ColBERT multi-vector embeddings support late-interaction search. Export tables, or render publication-ready plots with dark mode, glow, and custom fonts. A read-only mode and a Hugging Face downloader help you share finished scopes. Running on a dedicated VPS on RepoCloud with guaranteed CPU, RAM, and SSD, full root SSH access, and a browser serial console. MIT licensed.
Benefits
- Understand Data At A Glance
- Clusters and a 2D map reveal themes, outliers, and duplicates across thousands of rows, so you spot patterns that sorting and filtering a spreadsheet never shows.
- Every Step Stays Reproducible
- Each embedding, projection, and clustering run is saved as flat files with its parameters, so you can return to a promising experiment and repeat it exactly.
- Use Local Or Hosted Models
- Pick embedding and labeling models that run on your own machine or connect OpenAI, Cohere, Voyage, Mistral, or compatible endpoints, depending on privacy and budget.
- Share Findings With Teammates
- Publish scopes from a read-only server, export data and plots, or download shared scopes from Hugging Face, so colleagues explore the same map you built.
Features
- Guided Five-Step Setup
- Embed, project, cluster, label, and scope your data from one page, with a saved card for every run and a preview of the output.
- Interactive Explore View
- Search by meaning, browse cluster lists, select points on the map, and read full rows in a side panel linked to the visualization.
- Cluster Labeling Options
- Label clusters with top words or summaries from chat models, add dataset context to the prompt, and keep several label sets for one clustering.
- Image And ColBERT Support
- Embed images with CLIP-style models and browse a zoomable image map, or use multi-vector ColBERT embeddings for late-interaction similarity search.
- Plot And Data Export
- Download processed tables or generate static plots with adjustable labels, glow, kernel, resolution, and dark mode for reports and papers.
- Experiment Gallery
- Name and describe projection and clustering runs, then browse thumbnails with cluster quality metrics to choose the best settings.