CITATION.cff

cff-version: 1.2.0
title: "EmilsRecallIndex: Generative-Retrieval Memory Router vs. Embedding Retrieval on a Small Linked Franglais Personal-Memory Corpus"
message: "If you use this work, please cite it."
type: software
authors:
  - family-names: Devauchelle
    given-names: Emilien
    alias: emil1326
repository-code: "https://github.com/emil1326/EmilsRecallIndexModelTest"
abstract: >-
  A reproducible experiment testing whether a small LLM fine-tuned as a generative retriever
  (Differentiable Search Index; query -> note slug) beats a strong embedding-retrieval baseline
  (multilingual-e5-small + BM25) on a small, densely-linked, franglais personal-memory corpus,
  with emphasis on associative (two-hop) and multi-answer recall. The router is swept at 0.5B-3B
  (testing whether recall plateaus with size), composed with query decomposition, and contrasted with a
  from-scratch character-level CNN control that isolates whether the associative advantage is the routing
  formulation or pretrained semantics. Includes a hardened evaluation protocol (LLM-generated symptom-side
  and independent second-hop held-out queries, source-note exclusion, zero train/test leakage) and
  paired-bootstrap significance tests across all four arms.
keywords:
  - generative retrieval
  - differentiable search index
  - DSI
  - personal memory
  - associative retrieval
  - embedding retrieval
  - LoRA
  - franglais
license: CC-BY-NC-4.0
date-released: 2026-07-04
version: "1.2-routers-0.5B-1.5B-3B+charcnn"