Not In Our Time: Transport-Based Content Selection\\ as an Inspectable, Manipulable Alternative to RAG

Chris Brew

Keywords: optimal transport, min-cost flow, RAG, LLM stereotypes, editorial decision-making

TL;DR: A min-cost flow alternative to RAG that makes every passage-selection decision inspectable and manipulable, demonstrated on LLM-generated literary panel discussions.

Abstract:

We replace the embedding-similarity retrieval step in Retrieval-Augmented Generation with a min-cost flow formulation that makes every passage-selection decision visible as a cost, a flow value, and a constraint. The system enriches 6,916 passages from Dickens' Bleak House with structured LLM-generated metadata (20 fields covering narrative function, thematic weight, and emotional register),
then solves two transport problems: the first assigns passages to fictional expert personas whose demand profiles are encoded as integer vectors over provision dimensions, and the second allocates them to episode segments. The expert personas are, deliberately, stereotypes: a Marxist always finds class struggle; an audiobook actor always finds comedy; a close reader always finds prose rhythm.

Because the flow solver is cheap relative to LLM generation, users can rapidly explore alternative configurations — swapping panellists, doubling a character arc,
reweighting a dimension — and observe how each change reshapes both the passage selection (mean pairwise Jaccard similarity $\approx 0.52$ across twenty configurations) and the downstream vocabulary of the generated multi-voice dialogue.
Editorial decisions, encoded as demand vectors rather than implicit in a RAG prompt, become first-class objects hat can be inspected, compared, and argued about.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 147