A digest across a batch of documents
Each text is condensed on its own, and only the condensed versions are drawn together.
What the scenario looks like
- Execute + Data
- LLM
Handing ten documents to a model in one lump gets you a summary of the first and the last: attention is not spread evenly and the middle drops out. Condensing one at a time removes that entirely.
What is inside
- A list of texts — the entry, an array of strings.
- For each text — the loop. Its body condenses one document into three sentences.
- The overall digest receives only those condensations and turns them into a picture: what is shared, what differs, what to look at first.
Order is preserved: the loop collects results in the order it handed the pieces out.
The quietest failure here is a port name: the item reaches the model node on its Message
port, not “Input”. The prompt says {{ inputs.message }}; write inputs.input and the digest
is built out of three blanks, on a green run.
What to set up for yourself
- The condensing prompt. Currently three sentences: what it is about, what is proposed, what is asked of the reader. Emails want one thing, contracts another.
- The digest prompt. It asks not to list the condensations again — otherwise the result is just them glued together.
- Where the texts come from. They arrive in the message today. In practice their place is taken by reading mail, a collection query or a list of files.
What to connect
Nothing. Cost grows with the number of documents: one call per text plus one for the digest. Ten short documents is eleven turns, and that is still cheaper than reading them yourself.