feat(datasets): suggest long description from uploaded files

components/Datasets/DatasetLongDescriptionSuggestFromFiles.vue

@@ -0,0 +1,138 @@
+<template>
+  <div>
+    <div class="flex flex-wrap items-center gap-4 mt-2 mb-3">
+      <Tooltip v-if="!canSuggest">
+        <BrandedButton
+          type="button"
+          color="primary"
+          :disabled="true"
+        >
+          <div class="flex items-center space-x-2">
+            <RiSparklingLine
+              class="size-4"
+              aria-hidden="true"
+            />
+            <span>{{ $t('Suggérer une description à partir des fichiers') }}</span>
+          </div>
+        </BrandedButton>
+        <template #tooltip>
+          {{ tooltipWhenDisabled }}
+        </template>
+      </Tooltip>
+      <BrandedButton
+        v-else
+        type="button"
+        color="primary"
+        :icon="RiSparklingLine"
+        :loading="isGenerating"
+        @click="handleSuggest"
+      >
+        <template v-if="isGenerating">
+          {{ $t('Suggestion en cours…') }}
+        </template>
+        <template v-else>
+          {{ $t('Suggérer une description à partir des fichiers') }}
+        </template>
+      </BrandedButton>
+      <CdataLink
+        v-if="config.public.generateDescriptionFeedbackUrl && hasReceivedSuggestion"
+        :to="config.public.generateDescriptionFeedbackUrl"
+        target="_blank"
+        class="text-sm text-gray-medium"
+      >
+        {{ $t('Comment avez-vous trouvé cette suggestion ?') }}
+      </CdataLink>
+    </div>
+    <SimpleBanner
+      v-if="errorMessage"
+      type="danger"
+      class="mb-3"
+    >
+      {{ errorMessage }}
+    </SimpleBanner>
+  </div>
+</template>
+
+<script setup lang="ts">
+import { BrandedButton, Tooltip, SimpleBanner } from '@datagouv/components-next'
+import { RiSparklingLine } from '@remixicon/vue'
+import CdataLink from '~/components/CdataLink.vue'
+import type { ResourceForm } from '~/types/types'
+import { buildCombinedExcerptFromResourceForms } from '~/utils/read-dataset-file-excerpt'
+
+const description = defineModel<string>({ required: true })
+
+const props = defineProps<{
+  resources: Array<ResourceForm>
+  title: string
+  organization?: string
+}>()
+
+const { t } = useTranslation()
+const config = useRuntimeConfig()
+
+const isGenerating = ref(false)
+const hasReceivedSuggestion = ref(false)
+const errorMessage = ref<string | null>(null)
+
+const hasTitle = computed(() => !!props.title?.trim())
+const hasLocalFile = computed(() =>
+  props.resources.some(r => r.filetype === 'file' && !!r.file?.raw),
+)
+
+const canSuggest = computed(() => hasTitle.value && hasLocalFile.value)
+
+const tooltipWhenDisabled = computed(() => {
+  if (!hasTitle.value) {
+    return t('Renseignez d’abord le titre du jeu de données (étape précédente).')
+  }
+  if (!hasLocalFile.value) {
+    return t('Ajoutez au moins un fichier téléversé au format texte (CSV, JSON, etc.) pour utiliser cette fonctionnalité.')
+  }
+  return ''
+})
+
+const GENERIC_ERROR = 'Une erreur s’est produite lors de la suggestion. Vérifiez qu’au moins un fichier contient du texte lisible, puis réessayez.'
+
+async function handleSuggest() {
+  errorMessage.value = null
+  if (!canSuggest.value) {
+    return
+  }
+
+  try {
+    isGenerating.value = true
+    const excerpt = await buildCombinedExcerptFromResourceForms(props.resources)
+    if (!excerpt) {
+      errorMessage.value = t('Aucun extrait lisible n’a pu être lu depuis vos fichiers. Utilisez un fichier texte ou CSV avec des en-têtes, ou un fichier JSON/XML.')
+      return
+    }
+
+    const response = await $fetch<{ description?: string }>('/nuxt-api/albert/generate-dataset-long-description', {
+      method: 'POST',
+      body: {
+        title: props.title.trim(),
+        fileExcerpt: excerpt,
+        ...(props.organization?.trim() && { organization: props.organization.trim() }),
+      },
+    })
+
+    if (response.description?.trim()) {
+      description.value = response.description.trim()
+      hasReceivedSuggestion.value = true
+    }
+  }
+  catch (error) {
+    console.error('Failed to generate long dataset description:', error)
+    const raw = error && typeof error === 'object' && 'data' in error && error.data && typeof error.data === 'object' && 'statusMessage' in error.data
+      ? String((error.data as { statusMessage: string }).statusMessage)
+      : error instanceof Error
+        ? error.message
+        : ''
+    errorMessage.value = raw || GENERIC_ERROR
+  }
+  finally {
+    isGenerating.value = false
+  }
+}
+</script>

components/Datasets/New/Step3AddResources.vue

@@ -137,6 +137,13 @@
             </template>
           </LinkedToAccordion>
         </fieldset>
+        <DatasetLongDescriptionSuggestFromFiles
+          v-if="form.resources.length > 0"
+          v-model="datasetForm.description"
+          :resources="form.resources"
+          :title="datasetForm.title"
+          :organization="datasetForm.owned?.organization?.name"
+        />
         <Alert
           v-if="errors.length"
           type="error"
@@ -179,13 +186,15 @@
 
 <script setup lang="ts">
 import { BrandedButton, PaddedContainer, SimpleBanner } from '@datagouv/components-next'
+import DatasetLongDescriptionSuggestFromFiles from '../DatasetLongDescriptionSuggestFromFiles.vue'
 import UploadResourceModal from '../UploadResourceModal.vue'
 import type { DatasetForm, ResourceForm } from '~/types/types'
 
+const datasetForm = defineModel<DatasetForm>('datasetForm', { required: true })
+
 const props = defineProps<{
   loading: boolean
   resources: Array<ResourceForm>
-  datasetForm: DatasetForm
 }>()
 
 const emit = defineEmits<{
@@ -200,7 +209,7 @@ const { t } = useTranslation()
 const publishFileAccordionId = useId()
 const addDescriptionAccordionId = useId()
 
-const isDatasetOpen = computed(() => props.datasetForm.access_type === 'open')
+const isDatasetOpen = computed(() => datasetForm.value.access_type === 'open')
 
 const { form, getFirstError, getFirstWarning, touch, validate, errorsAsList: errors } = useForm({
   resources: props.resources,

components/Datasets/Structured/Step3DescribeDataset.vue

@@ -107,6 +107,13 @@
               :placeholder="$t('Décrivez le contenu, la source et l\'utilité de ces données...')"
               :rows="10"
             />
+            <DatasetLongDescriptionSuggestFromFiles
+              v-if="props.resources.length > 0"
+              v-model="form.description"
+              :resources="props.resources"
+              :title="form.title"
+              :organization="datasetForm.owned?.organization?.name"
+            />
           </LinkedToAccordion>
 
           <LinkedToAccordion
@@ -171,6 +178,7 @@
 import { BrandedButton, PaddedContainer, SimpleBanner, SearchableSelect } from '@datagouv/components-next'
 import type { Frequency } from '@datagouv/components-next'
 import { ref } from 'vue'
+import DatasetLongDescriptionSuggestFromFiles from '~/components/Datasets/DatasetLongDescriptionSuggestFromFiles.vue'
 import Alert from '~/components/Alert/Alert.vue'
 import InputGroup from '~/components/InputGroup/InputGroup.vue'
 import Sidemenu from '~/components/Sidemenu/Sidemenu.global.vue'
@@ -179,7 +187,13 @@ import AccordionGroup from '~/components/Accordion/AccordionGroup.global.vue'
 import LinkedToAccordion from '~/components/LinkedToAccordion/LinkedToAccordion.vue'
 import RequiredExplanation from '~/components/RequiredExplanation/RequiredExplanation.vue'
 import { useForm, required, minLength } from '~/composables/useForm'
-import type { DatasetForm } from '~/types/types'
+import type { DatasetForm, ResourceForm } from '~/types/types'
+
+const props = withDefaults(defineProps<{
+  resources?: Array<ResourceForm>
+}>(), {
+  resources: () => [],
+})
 
 const emit = defineEmits<{
   (e: 'previous' | 'next'): void

pages/admin/datasets/new.vue

@@ -33,8 +33,8 @@
     />
     <Step3AddResources
       v-if="currentStep === 3"
+      v-model:dataset-form="datasetForm"
       :resources
-      :dataset-form
       :loading
       @previous="moveToStep(2)"
       @next="filesNext"

pages/admin/datasets/structured.vue

@@ -45,6 +45,7 @@
     <Step3DescribeDataset
       v-if="currentStep === 3 && associateSchemaForm.selectedSchema && associateSchemaForm.owned?.organization"
       v-model="datasetForm"
+      :resources="resources"
       @previous="goBackFromStep3"
       @next="describeNext"
     />

server/routes/nuxt-api/albert/generate-dataset-long-description.post.ts

@@ -0,0 +1,89 @@
+import { DESCRIPTION_MIN_LENGTH } from '~/datagouv-components/src/functions/description'
+import { callAlbertAPI } from './utils/albert-helpers'
+import {
+  MAX_COMBINED_EXCERPT_CHARS,
+  MIN_COMBINED_EXCERPT_CHARS,
+} from '~/utils/read-dataset-file-excerpt'
+
+const MAXIMUM_PROMPT_LENGTH = 120_000
+
+export default defineEventHandler(async (event) => {
+  const body = await readBody(event)
+  const { title, organization, fileExcerpt } = body
+
+  if (!title?.trim()) {
+    throw createError({
+      statusCode: 400,
+      statusMessage: 'Title is required',
+    })
+  }
+
+  const excerpt = typeof fileExcerpt === 'string' ? fileExcerpt.trim() : ''
+  if (!excerpt || excerpt.length < MIN_COMBINED_EXCERPT_CHARS) {
+    throw createError({
+      statusCode: 400,
+      statusMessage: 'File excerpt is required and must contain enough text',
+    })
+  }
+
+  if (excerpt.length > MAX_COMBINED_EXCERPT_CHARS) {
+    throw createError({
+      statusCode: 400,
+      statusMessage: 'File excerpt exceeds maximum length',
+    })
+  }
+
+  const systemContent = `You are an assistant integrated into data.gouv.fr, the French open data platform.
+Your purpose is to help data producers write clear, comprehensive, and factual long descriptions of datasets.
+
+Guidelines:
+- Always respond in French.
+- Your tone is factual, neutral, and accessible to non-experts.
+- Use plain language and clear sentences; use Markdown when it helps structure (headings ##, bullet lists).
+- Do not make assumptions or add information that is not clearly supported by the excerpt or title.
+- Cover content, structure, and limits of the data when the excerpt allows it (variables, scope, updates, methodology if present).
+- Always start the main text with a capital letter.
+- IMPORTANT: Return ONLY the description text (Markdown allowed), without a preamble or labels such as "Description:".`
+
+  const userContent = `You are asked to generate a long description for a dataset on data.gouv.fr.
+
+Goal:
+→ Write a detailed, reusable description that helps people understand what the dataset contains and how to use it.
+→ Reflect only what can be inferred from the excerpt below (column names, codes, dates, geography, etc. when visible).
+→ If the excerpt is only a data sample, describe the likely subject matter and structure without inventing methodology or sources not shown.
+
+Dataset title: ${title.trim()}
+${organization ? `Producer organization: ${organization}\n` : ''}
+Excerpt from uploaded file(s):
+
+${excerpt}
+
+Output:
+→ A description in French with Markdown allowed (e.g. ## sections, lists).
+→ Minimum length: at least ${DESCRIPTION_MIN_LENGTH} characters.
+→ No generic filler about "open data" unless the excerpt supports it.`
+
+  const totalLength = systemContent.length + userContent.length
+  if (totalLength > MAXIMUM_PROMPT_LENGTH) {
+    throw createError({
+      statusCode: 422,
+      statusMessage: `The excerpt is too long to process (${totalLength} characters, maximum ${MAXIMUM_PROMPT_LENGTH}).`,
+    })
+  }
+
+  const messages = [
+    { role: 'system', content: systemContent },
+    { role: 'user', content: userContent },
+  ]
+
+  const generatedDescription = (await callAlbertAPI(messages, 'openweight-small')).trim()
+
+  if (generatedDescription.length < DESCRIPTION_MIN_LENGTH) {
+    throw createError({
+      statusCode: 422,
+      statusMessage: 'The model could not generate a sufficient description. Try a more descriptive text or CSV with headers.',
+    })
+  }
+
+  return { description: generatedDescription }
+})