Think · long documents · updated 22 October 2026

Digest

Shorten long text, a PDF or a scan faithfully, or pull named fields out of it. Not a chat: it does one job on what you give it.

See the recipes Models for /digest
  • 1credit for a short paste
  • 50credits reserved for the largest document
  • 20 MBlargest file
  • 1000summary tokens at most

Entry I

Start here

/digest summarize rewrites a long text in fewer words. /digest extract finds the fields you name. /digest sample shows filled-in examples (Summarize notes, Summarize a PDF, Extract invoice fields, Extract a receipt photo and Extract + yes/no field) and costs nothing.

Give the input as pasted message text (up to 4000 characters), a file (PDF, JPEG, PNG, TIFF or a text file, up to 20 MB), a public url to one, or a mix. Scans and photos are read with OCR first.

The price is reserved from the size of the input, from 1 credit up to 50 for a document filling the 1 million token window, and what is not used comes back.

Entry II

Summarize

Use instructions to set the format, language, focus or length: "Three bullet points. Keep every owner, date and decision." They add to the default rules; the summary stays faithful to the source. max_tokens caps the length (up to 1000). There is no temperature: sampling is fixed for faithful summaries.

Entry III

Extract

One field per line

Write name: what to find, one per line, in entities (up to 4000 characters). JSON works too.

Describe like an instruction

"total: Total amount due including tax, with currency" finds more than "total".

Lowercase names

Lowercase or Title Case field names work better than ALL CAPS.

Do not overlap

Asking for "location" and "city" in one run splits the hits between them. Pick one level.

Yes/no fields

A field whose value is a JSON object with labels is classified instead of copied; give each label a short rubric. /digest sample has a ready example.

Tune the output

threshold (0–1) drops weak hits, top_n (up to 50) caps hits per field, context_chars (up to 4000) sets the text shown around each hit. Every hit comes with a confidence.

Entry IV

Every option

What you can type after /digest, read from the command itself. You can run it again after 5 seconds.

/digest summarize Faithful rewrite of long text or a PDF/scan

OptionWhat it doesTakes
message Text to summarize (or a short note if you attach a file) textup to 4,000 characters
file PDF, any picture, or a text file (Discord upload cap) file
url Public http(s) link to a PDF, scan, or text file textup to 500 characters
instructions Format, language, focus, or length (adds to the default rules) textup to 1,000 characters
max_tokens Max summary tokens (default 1000, hard cap 1000) whole number1 to 1000
hidden Only you can see this (default: false — post in channel) true / false

/digest extract Pull named fields from text, a PDF, or a scan

OptionWhat it doesTakes
entities needed Fields: one `name: what to find` per line, or JSON textup to 4,000 characters
message Text to search (or a short note if you attach a file) textup to 4,000 characters
file PDF, any picture, or a text file (Discord upload cap) file
url Public http(s) link to a PDF, scan, or text file textup to 500 characters
instruction Rules for every field (dedupe, ranking). Not the field list. textup to 1,000 characters
threshold Drop hits below this confidence (0–1, default 0 = keep all) number0 to 1
top_n Max hits per field (0 = all, default 0) whole number0 to 50
context_chars Characters of context each side of a hit (default 500) whole number0 to 4000
hidden Only you can see this (default: false — post in channel) true / false

/digest sample Copy-ready examples for summarize and extract

OptionWhat it doesTakes
kind needed Which example to show textSummarize notes, Summarize a PDF, Extract invoice fields, Extract a receipt photo, Extract + yes/no field

Entry V

Recipes

Requests you can copy into Discord and change to fit.

Meeting notes in three bullets

Copy this

/digest summarize message: (paste the notes) instructions: Use three bullet points. Keep every owner, date and decision.

Why it works. Instructions set the shape; the facts stay.

Invoice fields from a PDF

Copy this

/digest extract entities: vendor: Name of the company issuing the invoice invoice_number: Invoice number or reference total: Total amount due including tax, with currency file: invoice.pdf

Why it works. Each field says in plain words what to look for.

Receipt photo

Copy this

/digest extract entities: merchant: Name of the store total: Total paid, with currency date: Date of the purchase file: receipt.jpg

Why it works. OCR reads the photo first; the raw text comes back with the fields.

Entry VI

Tips

  1. I

    Run /digest sample and copy the values into the real command.

  2. II

    Summaries of very long documents are better with a focus: "only the risks and the deadlines".

  3. III

    If a field comes back empty, try a synonym in its description.

Entry VII

Limits

  • Files up to 20 MB: PDF, JPEG, PNG, TIFF, or text (.txt, .md, .csv, .json, .html).

  • Links must be public http(s) files; private or local addresses are refused.

  • A summary is at most 1000 tokens.

  • 5 seconds between /digest commands.

Entry VIII

Questions

Why was I charged less than the reserve?

The reserve is an estimate from the input size. The real usage is charged and the rest refunded.

Can I ask questions about the document?

Not here. Digest summarizes or extracts. Paste the text into /chat to discuss it.

Does it translate?

Ask for a language in the summarize instructions; the summary is written in it.

/chatChat /spotSpot

Where the advice comes from

Prompting advice is from these free public guides, matched to what ALMA lets you set. Prices and limits are read from the bot itself, as of 22 October 2026.

Try it

Type /digest in any server with ALMA, or in a direct message to the bot.

All guides

Something here wrong, or missing? Tell us.