Text stack

The text stack follows the spatial design with one repository per role, so documents become tables an agent queries like any other layer.
Spatial stack
in production
Text stack
new
Recipes
geo-workflows
One reviewed recipe per dataset, from issue to published collection.
text-workflowspilot
One recipe per corpus, setting its sources, scope, languages and gazetteers.
Builder
geo-ingest
Converts vector and raster sources and indexes them on the H3 grid, as parallel Jobs.
text-ingestin design
Turns documents into tables of passages and references, with OCR, structure and places.
Data on S3
GeoParquetPMTilesCOGH3 hex parquet
documentspassagesreferencesH3 places
Metadata
A STAC catalog that the apps and the agent both read.
Croissant and CSL-JSON, with a thin STAC record so the catalog can find the corpus.
Query tool
geo-query
An MCP server for SQL over hex parquet, guided by STAC.
text-querylive in one app
An MCP server that runs codebooks over any text column a SQL query returns.
Running today in the private High Seas app
The app's agent codes passages from a literature corpus that was ingested by hand. The corpus sits in a private rustfs bucket, read with a key scoped to it.
high-seas-private high-seas-text-query high-seas-lit
One grid for both
Places named in a passage are stored as H3 cells, so coded passages join the spatial layers on the same grid.