feat: add summarization to pipeline (summarize + summary_sentences + summary_language)

server/main.py

@@ -63,6 +63,9 @@ class PipelineIngestRequest(BaseModel):
     generate_html: bool = True
     translate: bool = True
     target_language: str = "ko"
+    summarize: bool = False
+    summary_sentences: int = 5
+    summary_language: str | None = None
 
 
 @app.get("/health")
@@ -171,6 +174,9 @@ def pipeline_ingest(req: PipelineIngestRequest, _: None = Depends(require_api_ke
         generate_html=req.generate_html,
         translate=req.translate,
         target_language=req.target_language,
+        summarize=req.summarize,
+        summary_sentences=req.summary_sentences,
+        summary_language=req.summary_language,
     )
     return {"status": "ok", "doc_id": result.doc_id, "added": result.added_chunks, "chunks": result.chunks, "html_path": result.html_path}
 

server/pipeline.py

@@ -25,6 +25,38 @@ class DocumentPipeline:
         self.output_dir = Path(output_dir)
         (self.output_dir / "html").mkdir(parents=True, exist_ok=True)
 
+    def summarize(self, parts: List[str], target_language: str = "ko", sentences: int = 5) -> List[str]:
+        summarized: List[str] = []
+        sys_prompt = (
+            "당신은 전문 요약가입니다. 핵심 내용만 간결하게 요약하세요."
+        )
+        for p in parts:
+            if not p.strip():
+                summarized.append("")
+                continue
+            messages = [
+                {"role": "system", "content": sys_prompt},
+                {"role": "user", "content": (
+                    f"다음 텍스트를 {target_language}로 {sentences}문장 이내로 핵심만 요약하세요. 불필요한 수식어는 제거하고, 중요한 수치/용어는 보존하세요.\n\n{p}"
+                )},
+            ]
+            resp = self.ollama.chat(self.boost_model, messages, stream=False, options={"temperature": 0.2, "num_ctx": 32768})
+            content = resp.get("message", {}).get("content") or resp.get("response", "")
+            summarized.append(content.strip())
+        # 최종 통합 요약(선택): 각 청크 요약을 다시 결합해 더 짧게
+        joined = "\n\n".join(s for s in summarized if s)
+        if not joined.strip():
+            return summarized
+        messages2 = [
+            {"role": "system", "content": sys_prompt},
+            {"role": "user", "content": (
+                f"아래 부분 요약들을 {target_language}로 {max(3, sentences)}문장 이내로 다시 한번 통합 요약하세요.\n\n{joined}"
+            )},
+        ]
+        resp2 = self.ollama.chat(self.boost_model, messages2, stream=False, options={"temperature": 0.2, "num_ctx": 32768})
+        content2 = resp2.get("message", {}).get("content") or resp2.get("response", "")
+        return [content2.strip()]
+
     def translate(self, parts: List[str], target_language: str = "ko") -> List[str]:
         translated: List[str] = []
         sys_prompt = (
@@ -58,9 +90,30 @@ h1{{font-size: 1.6rem; margin-bottom: 1rem;}}
         html_path.write_text(html, encoding="utf-8")
         return str(html_path)
 
-    def process(self, *, doc_id: str, text: str, index, generate_html: bool = True, translate: bool = True, target_language: str = "ko") -> PipelineResult:
+    def process(
+        self,
+        *,
+        doc_id: str,
+        text: str,
+        index,
+        generate_html: bool = True,
+        translate: bool = True,
+        target_language: str = "ko",
+        summarize: bool = False,
+        summary_sentences: int = 5,
+        summary_language: str | None = None,
+    ) -> PipelineResult:
         parts = chunk_text(text, max_chars=1200, overlap=200)
-        translated = self.translate(parts, target_language=target_language) if translate else parts
+
+        if summarize:
+            # 요약 언어 기본값: 번역 언어와 동일, 번역 off면 ko로 요약(설정 없을 때)
+            sum_lang = summary_language or (target_language if translate else "ko")
+            summarized_parts = self.summarize(parts, target_language=sum_lang, sentences=summary_sentences)
+            working_parts = summarized_parts
+        else:
+            working_parts = parts
+
+        translated = self.translate(working_parts, target_language=target_language) if translate else working_parts
 
         to_append: List[IndexRow] = []
         for i, t in enumerate(translated):
@@ -70,7 +123,7 @@ h1{{font-size: 1.6rem; margin-bottom: 1rem;}}
 
         html_path: str | None = None
         if generate_html:
-            title_suffix = "번역본" if translate else "원문"
+            title_suffix = "요약+번역본" if (summarize and translate) else ("요약본" if summarize else ("번역본" if translate else "원문"))
             html_path = self.build_html(doc_id, title=f"문서 {doc_id} ({title_suffix})", ko_text="\n\n".join(translated))
 
         return PipelineResult(doc_id=doc_id, html_path=html_path, added_chunks=added, chunks=len(translated))