Source types

Mantis ships thirteen creation routes, all reachable from /new-space. They differ in what you supply, how much of the pipeline you control, and what each one is quietly limited by. This page is the reference; Create your first space is the walkthrough for the three most common paths.

The Guided mode of the create page splits these into Upload Documents and Gather Data. Cards mode shows all of them at once, and is the only mode that lists H5AD.

Upload what you have

AI Assistant

Route: /new-space/intelligent_synthesis
Takes: .csv, .xlsx

Upload a spreadsheet and let Mantis decide which columns are titles, which carry meaning, and which are metadata. You supply a space name, a map name, and the file. This is the same pipeline the home dashboard quick-create uses, and it is the fastest route if you do not want to think about column typing.

Use it when you are exploring a dataset for the first time, or the file has too many columns to classify by hand.

CSV/XLSX

Route: /new-space/tabular-upload
Takes: .csv, .xlsx, .json

The full tabular flow. You choose whether the map goes into a new space or an existing one, drop the file, and assign a type to every column: title, semantic, numeric, categoric, date, links, image, geospatial, coordinate1, coordinate2, vector, or unused. Space and map visibility are separate switches. Advanced processing settings exposes the embedding, reduction, clustering, and labeling configuration, generated live from the backend's schema.

Limits. Creation is blocked until the typing is valid: at least one semantic column, or a vector column paired with geospatial or a coordinate1/coordinate2 pair. Space and map names must be at least two characters.

Text

Route: /new-space/text
Takes: .txt, .pdf, .doc, .docx, or pasted text

For prose rather than tables. Upload multiple files (all of the same type) or paste text directly, then choose how the text is cut into points: one per chapter, one per page, one per section by heading, one per paragraph, one per sentence, or a custom segmentation prompt you write. You can preview the segmentation before committing.

Limits. Per-page segmentation is only available for PDF uploads. Mixed file types in one upload are rejected.

Audio

Route: /new-space/audio
Takes: MP3, WAV, M4A, up to 100 MB

Upload one or more recordings and get a space built from their content. Multi-file selection is supported.

Video

Route: /new-space/video
Takes: MP4, MOV, WebM, or a YouTube URL

Turns a lecture or talk into a semantic map of its transcript, with a Video Player and Transcript panel wired to the map. Choose Transcript Segmentation: by timestamp, scene, chapter, sentence, or a custom prompt. YouTube URLs are validated client-side against youtube.com and youtu.be.

Molecules

Route: /new-space/molecule_synthesis
Takes: .csv of SMILES strings

Lays molecules out by structural similarity using Morgan fingerprints. You set the fingerprint radius (1 to 4, default 2), the dimensionality reduction, and the clustering method, and can optionally compute molecular properties, enable clustering, and request GPU acceleration.

H5AD

Route: /new-space/h5ad
Takes: .h5ad / HDF5

Single-cell data. Mantis inspects the file, proposes a set of maps over one shared cell dataset, and lets you enable or disable each proposed map before materializing. Progress is reported per map. Creation requires a space name and at least one enabled map.

Only listed in Cards mode; the Guided flow does not offer it.

Gather data from elsewhere

Web Search

Route: /new-space/websearch
Takes: A query, or a list of URLs

Crawls the web and builds a space from what it finds. Search query mode takes a search term plus a Search Iterations count between 6 and 10, crawling ten websites per iteration. Enter specific URLs mode takes a list of pages, one per line.

Limits. A crawl producing 51 rows or fewer is rejected outright, with a message naming the row count it got.

Multimedia

Route: /new-space/MultiMedia
Takes: A query against a public API

A single interface over fourteen public data sources, grouped by category:

Knowledge & Reference: Wikipedia, SEC Filings Search
Scientific Research: PubMed, ArXiv, Figshare, Europe PMC
Earth, Life & Universe: NASA, GBIF, USGS Earthquakes, USGS Active Volcanoes
Cultural Heritage: Library of Congress, Open Library, Artic
Social Media: Reddit

Each source exposes its own parameter form (query, result limit, date ranges, sort order, and source-specific filters). The page also offers a Database Connection card for importing from your own database with a SQL query, and a natural-language search bar that returns a table you can turn into a space.

Academic

Route: /new-space/academic_synthesis
Papers: 50 to 5,000

Builds a space from an OpenAlex literature search. You give a query, optionally stack filters (Author, Institutions, Concepts, Open Access, Document Type, and more), and set the number of papers. A companion Search for OpenAlex ID panel resolves author, institution, and concept names to the IDs the filters expect; multiple IDs can be joined with & for AND or | for OR.

Limits. The paper count is clamped to the 50 to 5,000 range when you leave the field.

Academic Scraping

Route: /new-space/academic_scraping
Max PDFs: 1,000
Row floor: 50

A keyword-and-timeline scrape summarized by the Five Facets model. You give a research topic, a keyword list, a date window, and a maximum paper count of up to 1,000.

Limits. Only genuinely open-access PDFs are pulled, so you routinely get fewer files than the maximum you asked for. If the scrape yields fewer than 50 paragraph batches, ingestion stops and offers you the CSV to download instead, with a message telling you to widen the filters. Large batches take considerably longer.

GitHub Repo

Route: /new-space/github_synthesis
Takes: A repository URL

Maps a repository's files and structure. You supply the repository URL, a map name, an optional space name, and choose the embedding service and model, the chat completion service and model, and the reduction model directly on this form.

Periodically Updating

Route: /new-space/api_ingestion
Default interval: 5 minutes

Connects an external API and re-ingests it on a schedule, so the map keeps tracking the source. Configuration runs in two steps.

First, the connection: URL, HTTP method (GET or POST), request parameters and headers as JSON, an optional API key placed in a header or a query parameter, and a records path for digging into a nested response. Mantis then inspects the endpoint, shows you the columns and a sample, and suggests a type for each.

Second, the schedule and shape: pick a title column and a type per column (semantic, numeric, categoric, date, links, or delete), then set the ingestion interval on a slider that snaps between 1 second and 10 years, defaulting to 5 minutes. Advanced settings cover max points (default 5,000), max per ingest, rebuild every (0 means never recompute the whole projection, keeping clusters stable), and a dedupe key column (blank dedupes on full record content).

Limits. A semantic column is required. Choosing an aggressive interval against a rate-limited API is your problem, not the scheduler's.

Row counts

There is no global minimum or maximum number of rows. The real floors are local to specific paths:

Constraint	Where
3 observations	UMAP and PCA cannot run below this
50 rows	Academic scraping ingestion floor
50 papers	Academic search minimum
More than 51 rows	Web search crawl result
1,000 PDFs	Academic scraping ceiling
5,000 papers	Academic search ceiling
50 MB	Home dashboard quick-create file size
100 MB	Audio upload

Above those, spaces of millions of points are normal; the map renders them through a level-of-detail budget rather than refusing them.

Source types ​

Upload what you have ​

AI Assistant ​

CSV/XLSX ​

Text ​

Audio ​

Video ​

Molecules ​

H5AD ​

Gather data from elsewhere ​

Web Search ​

Multimedia ​

Academic ​

Academic Scraping ​

GitHub Repo ​

Periodically Updating ​

Row counts ​