feat: index the text of attached PDFs

- jz_bm_material_file_texts caches the pdftotext output per file
- the River job jz.bm.extract_material_text extracts and resyncs
- the management multipart attach queues it inside its transaction
- file_texts in the document and a file snippet with its name

models/material_file_text.go

@@ -0,0 +1,18 @@
+package models
+
+import "time"
+
+// MaterialFileText is the text extracted from one PDF attached to a
+// material. An empty Text means the file was extracted and held none (a
+// scan or a broken PDF). Truncated is set when the text was cut at the
+// extraction cap. The row goes with its file (foreign-key cascade).
+type MaterialFileText struct {
+	FileID      uint      `gorm:"column:file_id;primaryKey;autoIncrement:false"`
+	MaterialID  uint      `gorm:"column:material_id"`
+	Text        string    `gorm:"column:text"`
+	Truncated   bool      `gorm:"column:truncated"`
+	ExtractedAt time.Time `gorm:"column:extracted_at"`
+}
+
+// TableName is the file text table.
+func (MaterialFileText) TableName() string { return "jz_bm_material_file_texts" }

models/material_search.go

@@ -58,7 +58,71 @@ func (m *Material) ToSearchableArray(ctx context.Context, db *gorm.DB) (map[stri
 	if err != nil {
 		return nil, err
 	}
-	return MaterialDocument(*m, course, meetingNumber, tags, nil), nil
+	files, err := IndexedFiles(db, []uint{m.ID}, true)
+	if err != nil {
+		return nil, err
+	}
+	texts := make([]string, 0, len(files[m.ID]))
+	for _, f := range files[m.ID] {
+		texts = append(texts, f.Text)
+	}
+	return MaterialDocument(*m, course, meetingNumber, tags, texts), nil
+}
+
+// IndexedFile is one entry of a material document's file_texts: the file,
+// its name and, when asked for, its extracted text.
+type IndexedFile struct {
+	FileID   uint
+	FileName string
+	Text     string
+}
+
+// IndexedFiles returns, per material id, the files whose text the search
+// document holds, in document order: the current public files of the
+// material's files field that have a non-empty extracted text, by
+// sort_order (nulls last) and id. It is the one query that decides what
+// file_texts holds, so a snippet's array index maps back to a file through
+// it. A detached file is no longer a current file, so its text drops out
+// even before the cascade removes it. The text is read only when withText.
+func IndexedFiles(db *gorm.DB, materialIDs []uint, withText bool) (map[uint][]IndexedFile, error) {
+	out := map[uint][]IndexedFile{}
+	if len(materialIDs) == 0 {
+		return out, nil
+	}
+	ids := make([]string, 0, len(materialIDs))
+	for _, id := range materialIDs {
+		ids = append(ids, strconv.FormatUint(uint64(id), 10))
+	}
+	cols := "system_files.id AS file_id, system_files.file_name, system_files.attachment_id"
+	if withText {
+		cols += ", t.text"
+	}
+	var rows []struct {
+		FileID       uint
+		FileName     string
+		AttachmentID string
+		Text         string
+	}
+	err := db.Table("system_files").
+		Select(cols).
+		Joins("JOIN jz_bm_material_file_texts t ON t.file_id = system_files.id").
+		Where("system_files.attachment_type = ? AND system_files.field = ? AND system_files.is_public", Material{}.MorphName(), MaterialFilesField).
+		Where("system_files.attachment_id IN ?", ids).
+		Where("t.text <> ''").
+		Order("system_files.attachment_id, system_files.sort_order NULLS LAST, system_files.id").
+		Scan(&rows).Error
+	if err != nil {
+		return nil, err
+	}
+	for _, r := range rows {
+		n, err := strconv.ParseUint(r.AttachmentID, 10, 64)
+		if err != nil {
+			continue
+		}
+		id := uint(n)
+		out[id] = append(out[id], IndexedFile{FileID: r.FileID, FileName: r.FileName, Text: r.Text})
+	}
+	return out, nil
 }
 
 // MaterialDocument is the search document of m with its course, meeting

models/material_search_test.go

@@ -102,6 +102,16 @@ func TestMaterialDocument(t *testing.T) {
 		t.Errorf("file_texts = %v", doc["file_texts"])
 	}
 
+	// File texts keep their order (the IndexedFiles order); an empty list
+	// leaves the key out.
+	ordered := models.MaterialDocument(m, course, 3, tags, []string{"pierwszy plik", "drugi plik", "trzeci plik"})
+	if got, _ := ordered["file_texts"].([]string); !slices.Equal(got, []string{"pierwszy plik", "drugi plik", "trzeci plik"}) {
+		t.Errorf("file_texts = %v, want the three texts in order", ordered["file_texts"])
+	}
+	if _, ok := models.MaterialDocument(m, course, 3, tags, []string{})["file_texts"]; ok {
+		t.Error("an empty file text list carries file_texts")
+	}
+
 	// No code, no course id, no meeting, a draft, no files.
 	bare := models.MaterialDocument(models.Material{ID: 5, Title: "Szkic", Type: "inne"}, &models.Course{Name: "Teoria"}, 0, nil, nil)
 	if bare["course"] != "Teoria" {