feat(export): Volltranskript als Anhang in DOCX/PDF (meeting/phone-call/interview)

LLM komprimiert das Transkript bisher auf wenige Sätze + 20 Excerpt-Einträge
— der Rest fiel aus dem Dokument. Pipeline reicht Whisper-Segmente jetzt an
build_docx/build_pdf durch; Profile mit `docx.include_full_transcript: true`
hängen das komplette Transkript (9 pt) am Ende an.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

lxc-frontend/app/core/mac_client.py

@@ -146,16 +146,28 @@ class MacClient:
             job_id,
         )
 
-    async def export_docx(self, data: dict, profile: str | None = None, job_id: int | None = None) -> bytes:
+    async def export_docx(
+        self,
+        data: dict,
+        profile: str | None = None,
+        job_id: int | None = None,
+        full_transcript: list[dict] | None = None,
+    ) -> bytes:
         return await self._post_bytes_retry(
             f"{self.base_url}/api/export/docx",
-            {"data": data, "profile": profile},
+            {"data": data, "profile": profile, "full_transcript": full_transcript or []},
             job_id,
         )
 
-    async def export_pdf(self, data: dict, profile: str | None = None, job_id: int | None = None) -> bytes:
+    async def export_pdf(
+        self,
+        data: dict,
+        profile: str | None = None,
+        job_id: int | None = None,
+        full_transcript: list[dict] | None = None,
+    ) -> bytes:
         return await self._post_bytes_retry(
             f"{self.base_url}/api/export/pdf",
-            {"data": data, "profile": profile},
+            {"data": data, "profile": profile, "full_transcript": full_transcript or []},
             job_id,
         )

lxc-frontend/app/services/pipeline.py

@@ -166,6 +166,11 @@ async def run_pipeline(job_id: int) -> None:
             protocol_path.write_text(json.dumps(protocol, ensure_ascii=False, indent=2), encoding="utf-8")
             _update(job_id, protocol_path=str(protocol_path), progress=85)
 
+        # Volltranskript-Segmente: optionaler Anhang ans DOCX/PDF, falls das
+        # Profil `docx.include_full_transcript` setzt. Renderer ignoriert die
+        # Liste sonst — wir reichen sie trotzdem immer durch.
+        full_segments = tx.get("segments", []) or []
+
         # 4) DOCX
         if existing_dx_path:
             log.info("[job %s] reuse docx %s", job_id, existing_dx_path.name)
@@ -174,7 +179,8 @@ async def run_pipeline(job_id: int) -> None:
             log.info("[job %s] export docx", job_id)
             docx_bytes = await _with_heartbeat(
                 job_id,
-                client.export_docx(protocol, profile=profile, job_id=job_id),
+                client.export_docx(protocol, profile=profile, job_id=job_id,
+                                   full_transcript=full_segments),
             )
             docx_path = job_dir / "protocol.docx"
             docx_path.write_bytes(docx_bytes)
@@ -189,7 +195,8 @@ async def run_pipeline(job_id: int) -> None:
             log.info("[job %s] export pdf", job_id)
             pdf_bytes = await _with_heartbeat(
                 job_id,
-                client.export_pdf(protocol, profile=profile, job_id=job_id),
+                client.export_pdf(protocol, profile=profile, job_id=job_id,
+                                  full_transcript=full_segments),
             )
             pdf_path = job_dir / "protocol.pdf"
             pdf_path.write_bytes(pdf_bytes)

mac-worker/app/core/docx_export.py

@@ -19,7 +19,11 @@ from docx.shared import Pt
 from . import profiles
 
 
-def build_docx(data: dict[str, Any], profile_name: str | None = None) -> bytes:
+def build_docx(
+    data: dict[str, Any],
+    profile_name: str | None = None,
+    full_transcript: list[dict] | None = None,
+) -> bytes:
     profile = profiles.get(profile_name)
     cfg = profile.docx or {}
 
@@ -37,6 +41,9 @@ def build_docx(data: dict[str, Any], profile_name: str | None = None) -> bytes:
     for section in cfg.get("sections", []) or []:
         _render_section(doc, data, section)
 
+    if cfg.get("include_full_transcript") and full_transcript:
+        _render_full_transcript(doc, full_transcript)
+
     if data.get("_parse_error") and data.get("_raw"):
         doc.add_heading("Roh-Antwort des Modells", level=1)
         doc.add_paragraph(data["_raw"])
@@ -46,6 +53,23 @@ def build_docx(data: dict[str, Any], profile_name: str | None = None) -> bytes:
     return buf.getvalue()
 
 
+def _render_full_transcript(doc, segments: list[dict]) -> None:
+    """Hängt den Volltranskript-Anhang an: eine Überschrift, dann ein Absatz
+    pro Whisper-Segment in 9 pt (damit das Dokument nicht explodiert).
+    Leere Segmente werden übersprungen.
+    """
+    doc.add_page_break()
+    doc.add_heading("Volltranskript", level=1)
+    for seg in segments:
+        text = (seg.get("text") if isinstance(seg, dict) else str(seg)) or ""
+        text = text.strip()
+        if not text:
+            continue
+        p = doc.add_paragraph()
+        run = p.add_run(text)
+        run.font.size = Pt(9)
+
+
 def _render_section(doc, data: dict, section: dict) -> None:
     val = data.get(section["key"])
     if val in (None, "", [], {}):

mac-worker/app/core/pdf_export.py

@@ -69,7 +69,11 @@ def _para(text: str, style_name: str = "Body2") -> Paragraph:
     return Paragraph(_esc(text).replace("\n", "<br/>"), _STYLES[style_name])
 
 
-def build_pdf(data: dict[str, Any], profile_name: str | None = None) -> bytes:
+def build_pdf(
+    data: dict[str, Any],
+    profile_name: str | None = None,
+    full_transcript: list[dict] | None = None,
+) -> bytes:
     profile = profiles.get(profile_name)
     cfg = profile.docx or {}
 
@@ -108,6 +112,9 @@ def build_pdf(data: dict[str, Any], profile_name: str | None = None) -> bytes:
     for section in cfg.get("sections", []) or []:
         _render_section(story, data, section)
 
+    if cfg.get("include_full_transcript") and full_transcript:
+        _render_full_transcript(story, full_transcript)
+
     # Parse-Error-Fallback
     if data.get("_parse_error") and data.get("_raw"):
         story.append(Paragraph("Roh-Antwort des Modells", _STYLES["SectionHeading"]))
@@ -117,6 +124,27 @@ def build_pdf(data: dict[str, Any], profile_name: str | None = None) -> bytes:
     return buf.getvalue()
 
 
+def _render_full_transcript(story: list, segments: list[dict]) -> None:
+    """Anhang nach allen Sektionen: ein Absatz pro Segment in 9 pt.
+    Vor dem Block ein PageBreak, damit das Volltranskript auf einer
+    eigenen Seite beginnt und das Protokoll nicht zerreißt.
+    """
+    from reportlab.platypus import PageBreak  # local import — selten genutzt
+
+    story.append(PageBreak())
+    story.append(Paragraph("Volltranskript", _STYLES["SectionHeading"]))
+    transcript_style = ParagraphStyle(
+        name="TranscriptLine", parent=_STYLES["Body2"],
+        fontSize=9, leading=12, spaceAfter=3,
+    )
+    for seg in segments:
+        text = (seg.get("text") if isinstance(seg, dict) else str(seg)) or ""
+        text = text.strip()
+        if not text:
+            continue
+        story.append(Paragraph(_esc(text).replace("\n", "<br/>"), transcript_style))
+
+
 def _render_section(story: list, data: dict, section: dict) -> None:
     val = data.get(section["key"])
     if val in (None, "", [], {}):

mac-worker/app/main.py

@@ -82,11 +82,15 @@ class ProtocolIn(BaseModel):
 class DocxIn(BaseModel):
     data: dict
     profile: str | None = None
+    # Optional: Whisper-Segmente; werden vom Renderer nur genutzt, wenn das
+    # Profil `docx.include_full_transcript: true` setzt.
+    full_transcript: list[dict] = []
 
 
 class PdfIn(BaseModel):
     data: dict
     profile: str | None = None
+    full_transcript: list[dict] = []
 
 
 class PreloadIn(BaseModel):
@@ -170,7 +174,11 @@ async def protocol(payload: ProtocolIn):
 
 @app.post("/api/export/docx")
 async def export_docx(payload: DocxIn):
-    data = build_docx(payload.data, profile_name=payload.profile)
+    data = build_docx(
+        payload.data,
+        profile_name=payload.profile,
+        full_transcript=payload.full_transcript,
+    )
     return Response(
         content=data,
         media_type="application/vnd.openxmlformats-officedocument.wordprocessingml.document",
@@ -180,7 +188,11 @@ async def export_docx(payload: DocxIn):
 
 @app.post("/api/export/pdf")
 async def export_pdf(payload: PdfIn):
-    data = build_pdf(payload.data, profile_name=payload.profile)
+    data = build_pdf(
+        payload.data,
+        profile_name=payload.profile,
+        full_transcript=payload.full_transcript,
+    )
     return Response(
         content=data,
         media_type="application/pdf",

mac-worker/profiles/interview.yaml

@@ -44,6 +44,7 @@ protocol_prompt: |
 docx:
   title_field: title
   default_title: Bewerbergespräch
+  include_full_transcript: true
   meta:
     - { key: date,      label: "Datum" }
     - { key: candidate, label: "Bewerber/in" }

mac-worker/profiles/meeting.yaml

@@ -50,6 +50,10 @@ protocol_prompt: |
 docx:
   title_field: title
   default_title: Sitzungsprotokoll
+  # Hängt das vollständige Whisper-Transkript als Anhang ans Dokument.
+  # Pipeline reicht die Segmente an den Renderer durch; LLM-Komprimierung
+  # im transcript_excerpt-Feld bleibt unverändert.
+  include_full_transcript: true
   meta:
     - { key: date,  label: "Datum" }
     - { key: topic, label: "Thema" }

mac-worker/profiles/phone-call.yaml

@@ -45,6 +45,7 @@ protocol_prompt: |
 docx:
   title_field: title
   default_title: Telefonnotiz
+  include_full_transcript: true
   meta:
     - { key: date,    label: "Datum" }
     - { key: caller,  label: "Anrufer" }