Prompts · Humathèque POC catalogage

extraction_these.txt

Envoyé tel quel au VLM avec les images du document. Le schéma JSON (entre accolades) fixe les champs extraits ; la casse, les accents et les civilités sont pris en charge par le VLM.

You are extracting metadata from a French graduate thesis title page.
General rules:
- Return ONLY valid JSON.
- Do not add markdown.
- Do not add explanations.
- Do not invent missing data.
- Multivalued fields are always JSON arrays; use [] when empty. Scalar fields use null when absent.
- If source text is written in all-caps, transform it into sentence case with French accents.
- For person names, apply sentence case (e.g. Jean-Daniel DUBOIS → Jean-Daniel Dubois, MARIE-FRANCE Dupont → Marie-France Dupont) and remove all civility and title prefixes (e.g. Monsieur, Madame, M., Mme, Professeur, Pr, Dr, etc.) before returning the name.
Role attribution rules:
- A role is explicit only when a textual role label is directly associated with the person.
- Assign a person to a specific role ONLY if that role is explicitly written on the title page.
- Do NOT infer roles from position, ordering, typography, academic conventions, or prior knowledge.
- Do not infer role equivalence from document structure or visual grouping.
- If people are listed without explicit role labels, place them ONLY in committee_members.
- If no jury or committee members are explicitly listed, use [].
- Do NOT guess who is advisor, reviewer, or jury president.
{
  "title": "Main title as it appears on the title page",
  "subtitle": "Subtitle or remainder of the title, usually following a colon.",
  "author": "Full name of the author (student) who wrote the graduate thesis",
  "degree_type": "Academic degree sought by the author. Possible values are Thèse d'État | Thèse de doctorat | Thèse de 3e cycle | Thèse d'université | Thèse de docteur-ingénieur | Thèse d'exercice",
  "discipline": "Academic field or discipline of the graduate thesis as explicitly mentioned. If not explicitly mentioned, infer it from the title and any other visible contextual elements (this is the only field where inference is allowed). Always return the discipline label in French (e.g. Histoire, Droit public, Sciences de l'éducation).",
  "volume": "Volume number if the title page explicitly mentions a volume, returned as the Arabic numeral only (e.g. \"1\" when the page prints Tome 1, TOME PREMIER or vol. I). Null if not explicitly mentioned.",
  "granting_institution": "Institution where the graduate thesis was submitted and the degree is granted",
  "co_tutelle_institutions": "List of institutions involved in a joint supervision or co-tutelle agreement.",
  "doctoral_school": "Doctoral school or graduate program, if explicitly mentioned",
  "defense_year": "Year the graduate thesis was defended. Format yyyy",
  "advisor": "List of ONLY persons explicitly identified as graduate thesis advisor/supervisor/director.",
  "jury_president": "ONLY the person explicitly identified as president/chair of the jury. Null if not explicitly stated.",
  "reviewers": "List of ONLY official reviewers/rapporteurs explicitly identified as such.",
  "committee_members": "List of persons explicitly listed as jury or committee members whose role is not explicitly identified as president/chair or reviewer/rapporteur. Include unlabeled persons here.",
  "language": "Language in ISO 639-3 codes. Example: fre, eng, ita...",
  "confidence": "Confidence score between 0.0 and 1.0 reflecting certainty based only on explicitly visible evidence on the page"
}

Catalogueur

L'enregistrement vérifie que le schéma JSON reste lisible, archive la version précédente dans config/prompts/archives/ et journalise la modification.

Champ du prompt	Variable de la notice · transformation	Liste	Active
title	aucune (écrit par le VLM)
subtitle	première lettre en minuscule (sauf sigle)
author	personne · découpage nom / prénom ; PPN d'IdRef si aligné
degree_type	aucune
discipline	aucune (écrit par le VLM)
volume	aucune
granting_institution	organisme · PPN d'IdRef si aligné ; sert aussi au calcul ephe (pipeline)
co_tutelle_institutions	organisme · PPN d'IdRef si aligné
doctoral_school	organisme · PPN d'IdRef si aligné
defense_year	contrôle du format aaaa (sinon vide + alerte)
advisor	personne · découpage nom / prénom ; PPN d'IdRef si aligné
jury_president	personne · découpage nom / prénom ; PPN d'IdRef si aligné
reviewers	personne · découpage nom / prénom ; PPN d'IdRef si aligné
committee_members	personne · découpage nom / prénom ; PPN d'IdRef si aligné
language	aucune
confidence	—
abstract absent des prompts	aucune
abstract_lang absent des prompts	aucune
parallel_title absent des prompts	aucune
parallel_title_lang absent des prompts	aucune
partner_institutions absent des prompts	organisme · aucune

Variable	Source	Calcul
titre_classement	calcul	titre + @ devant le premier mot significatif (article initial selon la langue)
volumaison	calcul	« 1 volume » si volume est vide, « ? volumes » sinon
code_etab	constante	CODE_ETAB de .env
numero_nnt indisponible	calcul	Numéro incrémental du NNT (fichier des numéros à définir)
dewey	service.dewey-classifier	première classe retenue par dewey-classifier

Prompts d'extraction

extraction_these.txt

Du prompt à la notice

Variables calculées ou fournies par un service