Create your first space

Every space starts at /new-space, reached from New Space on the home dashboard or Ctrl+N. You pick a source type, tell Mantis which parts of your data carry meaning, and it embeds, projects, clusters, and labels the result. This page covers the three paths most people take first. Source types documents all thirteen.

Choose a source

The create page offers thirteen routes:

Source	Route	Use it for
AI Assistant	/new-space/intelligent_synthesis	A spreadsheet whose columns you would rather not classify by hand
CSV/XLSX	/new-space/tabular-upload	A spreadsheet where you want control over every column
Text	/new-space/text	TXT, PDF, DOC, DOCX, or pasted prose
Audio	/new-space/audio	Recordings, in MP3, WAV, MP4, or M4A
Video	/new-space/video	A lecture or talk, uploaded or from a YouTube URL
Web Search	/new-space/websearch	A search query to crawl, or a list of URLs
Multimedia	/new-space/MultiMedia	Wikipedia, PubMed, arXiv, NASA, Reddit and other public APIs
Academic	/new-space/academic_synthesis	An OpenAlex literature search with filters
Academic Scraping	/new-space/academic_scraping	A topic-and-keyword scrape summarized by Five Facets
GitHub Repo	/new-space/github_synthesis	A repository’s files and structure
Molecules	/new-space/molecule_synthesis	A CSV of SMILES strings
H5AD	/new-space/h5ad	Single-cell expression matrices and their metadata
Periodically Updating	/new-space/api_ingestion	An external API re-ingested on a schedule

The page opens in one of two modes, and remembers which one you used last:

• Guided asks whether you want to Upload Documents or Gather Data, then shows only the sources in that half, tagged easy, medium, or advanced.
• Cards shows every source at once. Use it if you already know what you want.

The H5AD path appears in Cards but not in Guided; go to /new-space/h5ad directly if you need it.

CSV or XLSX

The tabular path is the most direct: you choose the destination, drop a file, and tell Mantis what each column is.

1. Pick a destination. Create a new space, or add this map to an existing one. Both names must be at least two characters.
2. Drop the file. The dropzone takes .csv, .xlsx, and .json. The space and map names are pre-filled from the filename.
3. Type the columns. Mantis previews the first rows and guesses a type per column. Correct anything it got wrong.
4. Set visibility. Make Space Public and Make Map Public are independent switches, both off by default.
5. Click Create Space. You are redirected to the progress page for the new map while embedding and reduction run.

What the column types mean

Type	Meaning
title	The name or label shown for the point
semantic	Free text that drives the embedding. This is the one that decides the layout
numeric	Numbers, available to plots, gradients, and feature analysis
categoric	Groups and categories
date	Time values
links	External URLs
image	Image data, such as base64 strings
geospatial	Geographic values
coordinate1, coordinate2	Fixed x and y, bypassing the embedding entirely
vector	A precomputed numeric vector
unused	Excluded from the embedding

Create Space stays disabled until the typing is usable: you need at least one semantic column, or a vector column paired with either geospatial or a single coordinate1/coordinate2 pair. The form says so in place if the combination is not valid.

Column types available on the tabular upload path. This screenshot predates the current grid layout, but the types themselves are current.

Let AI type the columns

If you would rather not classify columns yourself, use AI Assistant (/new-space/intelligent_synthesis) instead. Same file, same result, but the column analysis is automatic. The quick-create dropzone on the home dashboard uses this path.

Text

Use this for documents rather than tables: research notes, transcripts, papers, or pasted prose.

1. Open the Text card (/new-space/text). Name the space and the map.
2. Choose an input method. Upload text documents (pdf, doc, docx, txt), which accepts multiple files of the same type, or Enter text manually, which takes pasted text.
3. Choose a segmentation mode. This decides what becomes one point: one per chapter (fastest), per page (PDF uploads only), per section by heading, per paragraph, per sentence, or a custom segmentation prompt you write yourself.
4. Preview and create. You can preview the segmentation before committing.

Segmentation is the important decision here. Per-sentence gives you a dense map of fine-grained claims; per-chapter gives you a sparse map of whole documents. You cannot change it after the fact without rebuilding the map.

Web Search

Build a space out of live web content rather than a file you already have.

1. Open the Web Search card (/new-space/websearch). Name the space and the map.
2. Pick an input method. Search query crawls the web from a search term. Enter specific URLs takes a list of pages, one per line.
3. Set the number of search iterations if you are searching. Ten websites are crawled per iteration, and the field accepts 6 to 10.
4. Create, and watch the crawling progress.

A crawl that returns 51 rows or fewer is rejected: the page reports how many rows it extracted and stops rather than building a map too sparse to project. Widen the query or add iterations and try again.

Creating a space from web search results. The form has changed since this screenshot was taken; the flow has not.

Advanced options

The tabular path hides its pipeline configuration behind Advanced processing settings, and the defaults are chosen so that you can ignore it entirely. The form is generated from the backend’s config schema, so what you see reflects what the server actually supports rather than a hard-coded list. It has four blocks:

• Embedding: method and service. The default is the Mantis vLLM embedding service with automatic model selection.
• Reduction: UMAP, TSNE, PCA, PCA+UMAP, PARAMETRIC_UMAP, or GAUSSIAN_UMAP. UMAP is the default, with 15 neighbors and a minimum distance of 0.1.
• Clustering: how points are grouped once they are laid out.
• Labeling: how those clusters get their names.

Whatever you pick, reduction outputs 2 dimensions. There is no 3D output to choose; 2.5D in the map is a camera mode.

Other creation paths expose their own subset: the GitHub and molecule forms ask for embedding, chat, and reduction directly, while the home dashboard quick-create asks for nothing at all.

Add another map to a space

A space can hold more than one map over the same records, which is how you compare different framings of one dataset.

• From the home dashboard. The Add New Map button on a space card, and New Map in the Manage Maps modal, both open the tabular upload flow pre-targeted at that space.
• From inside the space. Open the Space Creation panel and go to the Add Map tab. Turn on Alternative View, click Load Metadata Columns, tick the metadata columns the new map should be built from, and submit. This path derives a map from data already in the space rather than from a new upload.
• Through the API. Developer payloads can still mark a column connection: true to link points to a previous map. The field survives in the API and in the CSV column vocabulary; it is not offered in any current creation grid.

There is no row limit

Mantis does not enforce a minimum or maximum number of rows on the generic upload paths. Real floors exist where the source itself is the constraint: about 50 rows on academic scraping and academic search, more than 51 on web search, and 3 observations before UMAP or PCA can run at all. Above that there is no documented ceiling.

Read nextSource typesEvery creation route, what it needs, and what it actually limits.