Skip to main content
Tables become queryable sources. Documents remain searchable text read directly from the selected workspace; Fella does not create an embedding index.

Supported formats

PDF extraction handles text-layer PDFs. Scanned or photographed documents need OCR before Fella can read them.

Messy values

The ingestors infer integer, float, boolean, date, and text shapes. They can parse written amounts such as $1,200.50, treat blank-like values as missing in numeric columns, normalize unambiguous named-month dates, skip preamble rows, and drop a trailing total row. Caveats are supplied to the agent through table inspection and can appear in evidence.

Tables and names

Use /files to see which table a file became. Use /schema <name> for its row count, column names, types, and null percentages. During an agent run, inspect_table can additionally return distinct counts, min/max values, ingest notes, and sample rows.

Documents are direct reads

grep_files searches catalogued document text with a case-insensitive regular expression and returns file/line matches. read_file accepts one or several document names. Each document is capped at 12,000 characters and a multi-document call is capped at 16,000 characters.
PDF support reads selectable text only. Image-only scans return a no-text warning; Fella does not perform OCR.

Rescan after changes

/reindex refreshes the catalog. /files shows loaded sources and noticed files that could not be used. A root .fellaignore can exclude names and workspace-relative path prefixes.
fella.md at the workspace root is context, not a document source. Fella skips it during the data scan and injects it separately into the model prompt.

Open a folder

Set the workspace boundary and understand scanning.

Ask questions

Match a question to tables, documents, or analysis tools.