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

controllers/api/material_search.go

@@ -9,6 +9,7 @@ import (
 	"git.golem15.com/golem15/summercms/modules/beachcomber"
 	"git.golem15.com/golem15/summercms/modules/wire"
 	"git.golem15.com/jakub/sm-bm-plugin/models"
+	"gorm.io/gorm"
 )
 
 // Material search: the engine (beachcomber, Typesense) answers candidate ids
@@ -66,17 +67,20 @@ type snippetJSON struct {
 }
 
 // pickSnippet returns the snippet of m from hit: the first body snippet,
-// mapped one to one. It is nil when the hit has none. Call it only for a
-// row the SQL re-gate returned.
-func pickSnippet(hit beachcomber.Hit, m models.Material) *snippetJSON {
+// mapped one to one, or else the first file_texts snippet, with the name of
+// the file at its array index in files (the IndexedFiles of m, in document
+// order; nil when the index does not resolve). A file text is long and is
+// not loaded for the page, so a file snippet is always cut on both sides.
+// It is nil when the hit has neither. Call it only for a row the SQL
+// re-gate returned.
+func pickSnippet(hit beachcomber.Hit, m models.Material, files []models.IndexedFile) *snippetJSON {
 	for _, sn := range hit.Highlights {
 		if sn.Field != "body" || len(sn.Segments) == 0 {
 			continue
 		}
-		out := &snippetJSON{Source: "body", Segments: make([]segmentJSON, 0, len(sn.Segments))}
+		out := &snippetJSON{Source: "body", Segments: snippetSegments(sn)}
 		var joined strings.Builder
 		for _, seg := range sn.Segments {
-			out.Segments = append(out.Segments, segmentJSON{Text: seg.Text, Hit: seg.Match})
 			joined.WriteString(seg.Text)
 		}
 		body := ""
@@ -88,9 +92,66 @@ func pickSnippet(hit beachcomber.Hit, m models.Material) *snippetJSON {
 		out.Trail = !strings.HasSuffix(body, text)
 		return out
 	}
+	for _, sn := range hit.Highlights {
+		if sn.Field != fileTextsField || len(sn.Segments) == 0 {
+			continue
+		}
+		out := &snippetJSON{Source: "file", Lead: true, Trail: true, Segments: snippetSegments(sn)}
+		if sn.ArrayIndex >= 0 && sn.ArrayIndex < len(files) {
+			name := files[sn.ArrayIndex].FileName
+			out.FileName = &name
+		}
+		return out
+	}
 	return nil
 }
 
+// fileTextsField is the document field of the attached files' texts.
+const fileTextsField = "file_texts"
+
+// snippetSegments maps the engine's segments one to one.
+func snippetSegments(sn beachcomber.Snippet) []segmentJSON {
+	out := make([]segmentJSON, 0, len(sn.Segments))
+	for _, seg := range sn.Segments {
+		out = append(out, segmentJSON{Text: seg.Text, Hit: seg.Match})
+	}
+	return out
+}
+
+// hasFileSnippet reports whether hit carries a file_texts snippet and no
+// body one, so its snippet needs the material's file names.
+func hasFileSnippet(hit beachcomber.Hit) bool {
+	file := false
+	for _, sn := range hit.Highlights {
+		if len(sn.Segments) == 0 {
+			continue
+		}
+		switch sn.Field {
+		case "body":
+			return false
+		case fileTextsField:
+			file = true
+		}
+	}
+	return file
+}
+
+// snippetFiles loads, once per page, the indexed files (names only) of the
+// rows whose snippet comes from a file. It reads nothing when no row needs
+// it. ids are rows SQL returned.
+func snippetFiles(db *gorm.DB, ids []uint, hits map[uint]beachcomber.Hit) (map[uint][]models.IndexedFile, error) {
+	var need []uint
+	for _, id := range ids {
+		if hit, ok := hits[id]; ok && hasFileSnippet(hit) {
+			need = append(need, id)
+		}
+	}
+	if len(need) == 0 {
+		return nil, nil
+	}
+	return models.IndexedFiles(db, need, false)
+}
+
 // flattenSnippet is a snippet as plain text, for MCP: the segment texts
 // joined, with an ellipsis where text was cut.
 func flattenSnippet(s *snippetJSON) string {