Reduce LLM rewrite calls and add report intro conclusion

ai_daily_report/assemble.py

@@ -35,13 +35,28 @@ def _source_link(item: NewsItem) -> str:
     return source
 
 
+def _fallback_intro(items: list[NewsItem]) -> str:
+    count = len(items)
+    return f"今天共聚合 {count} 条 AI 动态，覆盖模型能力、产品应用、基础设施、资本与治理等方向。"
+
+
+def _fallback_conclusion(items: list[NewsItem]) -> str:
+    sections = [section for section in SECTION_ORDER if any(item.section == section for item in items)]
+    if sections:
+        return "总体看，今日 AI 动态主要集中在" + "、".join(sections[:4]) + "等方向，后续仍需持续观察落地进展。"
+    return "总体看，今日 AI 动态仍在持续演进，后续需要关注产品落地和生态变化。"
+
+
 def assemble_markdown(items: list[NewsItem], guide: dict[str, Any] | None = None) -> tuple[str, dict[str, Any]]:
-    guide = guide or {"theme": "", "threads": []}
+    guide = guide or {"intro": "", "theme": "", "threads": [], "conclusion": ""}
     lines: list[str] = []
 
+    intro = _ensure_sentence(str(guide.get("intro") or "")) or _fallback_intro(items)
+    lines.extend(["## 引言", "", f"> {intro}", ""])
+
     theme = _clean_text(str(guide.get("theme") or ""))
     if theme:
-        lines.extend(["## 导览", "", f"> {theme}", ""])
+        lines.extend(["## 导览", "", f"> {_ensure_sentence(theme)}", ""])
 
     item_number = 1
     for section in SECTION_ORDER:
@@ -72,6 +87,9 @@ def assemble_markdown(items: list[NewsItem], guide: dict[str, Any] | None = None
                 continue
             lines.extend([f"- **{title}**", f"  {text}", ""])
 
+    conclusion = _ensure_sentence(str(guide.get("conclusion") or "")) or _fallback_conclusion(items)
+    lines.extend(["## 总结", "", f"> {conclusion}", ""])
+
     markdown = "\n".join(lines).strip()
     report = validate_markdown(markdown, items)
     return markdown, report

ai_daily_report/guide.py

@@ -23,8 +23,10 @@ def _clean_text(text: str, limit: int | None = None) -> str:
 def _build_prompt(items: list[NewsItem]) -> str:
     payload = {
         "task": (
-            "Generate a concise AI daily report guide. Return JSON only. Do not use 强信号/中信号/待验证. "
+            "Generate a concise Chinese AI daily report guide. Return JSON only. "
-            "Use a short theme and 2-4 daily threads. Every thread must reference existing item_ids."
+            "Do not use 强信号/中信号/待验证. Do not add facts. "
+            "Write one opening intro, a short theme, 2-4 daily threads, and one closing conclusion. "
+            "Every thread must reference existing item_ids."
         ),
         "items": [
             {
@@ -37,6 +39,7 @@ def _build_prompt(items: list[NewsItem]) -> str:
             for item in items
         ],
         "output_schema": {
+            "intro": "one opening paragraph under 160 Chinese characters",
             "theme": "one sentence under 120 Chinese characters",
             "threads": [
                 {
@@ -46,23 +49,27 @@ def _build_prompt(items: list[NewsItem]) -> str:
                     "kind": "thread|uncertain",
                 }
             ],
+            "conclusion": "one closing paragraph under 180 Chinese characters",
         },
     }
     return json.dumps(payload, ensure_ascii=False)
 
 
+def _empty_guide() -> dict[str, Any]:
+    return {"intro": "", "theme": "", "threads": [], "conclusion": ""}
+
+
 def generate_guide(
     items: list[NewsItem],
     *,
     llm_call: GuideLlmCall,
 ) -> tuple[dict[str, Any], dict[str, Any]]:
     if not items:
-        return {
+        return _empty_guide(), {
-            "theme": "",
-            "threads": [],
-        }, {
             "input_count": 0,
+            "intro_present": False,
             "theme_present": False,
+            "conclusion_present": False,
             "thread_count": 0,
             "dropped_thread_count": 0,
             "fallback_used": False,
@@ -72,12 +79,11 @@ def generate_guide(
     try:
         obj = parse_json_object(llm_call(_build_prompt(items)))
     except Exception as exc:
-        return {
+        return _empty_guide(), {
-            "theme": "",
-            "threads": [],
-        }, {
             "input_count": len(items),
+            "intro_present": False,
             "theme_present": False,
+            "conclusion_present": False,
             "thread_count": 0,
             "dropped_thread_count": 0,
             "fallback_used": True,
@@ -100,11 +106,15 @@ def generate_guide(
         kind = thread.get("kind") if thread.get("kind") in ("thread", "uncertain") else "thread"
         threads.append({"title": title, "text": text, "item_ids": item_ids, "kind": kind})
 
+    intro = _clean_text(str(obj.get("intro") or ""), limit=160)
     theme = _clean_text(str(obj.get("theme") or ""), limit=120)
-    guide = {"theme": theme, "threads": threads}
+    conclusion = _clean_text(str(obj.get("conclusion") or ""), limit=180)
+    guide = {"intro": intro, "theme": theme, "threads": threads, "conclusion": conclusion}
     report = {
         "input_count": len(items),
+        "intro_present": bool(intro),
         "theme_present": bool(theme),
+        "conclusion_present": bool(conclusion),
         "thread_count": len(threads),
         "dropped_thread_count": dropped,
         "fallback_used": False,

ai_daily_report/rewrite.py

@@ -83,8 +83,9 @@ def rewrite_items(
     items: list[NewsItem],
     *,
     llm_call: RewriteLlmCall,
-    batch_size: int = 10,
+    batch_size: int = 30,
     max_fallback_ratio: float = 0.2,
+    retry_single_items: bool = False,
 ) -> tuple[list[NewsItem], dict[str, Any]]:
     rewritten_count = 0
     fallback_count = 0
@@ -100,6 +101,11 @@ def rewrite_items(
                     _fallback(item)
                     fallback_count += 1
                 continue
+            if not retry_single_items:
+                for item in batch:
+                    _fallback(item)
+                    fallback_count += 1
+                continue
             for item in batch:
                 try:
                     rewritten_count += _apply_rewrite_batch([item], llm_call)

ai_daily_report/runner.py

@@ -64,6 +64,7 @@ def _mock_guide_llm(prompt: str) -> str:
     item_ids = [item["id"] for item in payload["items"][:3]]
     return json.dumps(
         {
+            "intro": "本地 mock 模式已生成 AI 日报，用于验证流水线。",
             "theme": "本地 mock 模式已生成 AI 日报，用于验证流水线。",
             "threads": [
                 {
@@ -73,6 +74,7 @@ def _mock_guide_llm(prompt: str) -> str:
                     "kind": "thread",
                 }
             ],
+            "conclusion": "本地 mock 结果可用于确认定时任务入口和文件输出是否正常。",
         },
         ensure_ascii=False,
     )

tests/test_stage4_rewrite.py

@@ -62,7 +62,7 @@ class Stage4RewriteTests(unittest.TestCase):
         self.assertEqual(report["fallback_count"], 1)
         self.assertIn("TimeoutError", report["errors"][0])
 
-    def test_rewrite_items_retries_failed_batch_as_single_items(self):
+    def test_rewrite_items_can_retry_failed_batch_as_single_items_when_enabled(self):
         items = [news_item("a"), news_item("b")]
         calls = []
 
@@ -85,13 +85,53 @@ class Stage4RewriteTests(unittest.TestCase):
                 }
             )
 
-        rewritten, report = rewrite_items(items, llm_call=llm_call, batch_size=2)
+        rewritten, report = rewrite_items(items, llm_call=llm_call, batch_size=2, retry_single_items=True)
 
         self.assertEqual([item.title for item in rewritten], ["title a", "title b"])
         self.assertEqual(report["rewritten_count"], 2)
         self.assertEqual(report["fallback_count"], 0)
         self.assertEqual(calls, [["a", "b"], ["a"], ["b"]])
 
+    def test_rewrite_items_does_not_retry_single_items_by_default(self):
+        items = [news_item("a"), news_item("b")]
+        calls = []
+
+        def llm_call(prompt):
+            payload = json.loads(prompt)
+            calls.append([item["id"] for item in payload["items"]])
+            return "not json"
+
+        rewritten, report = rewrite_items(items, llm_call=llm_call, batch_size=2)
+
+        self.assertEqual(calls, [["a", "b"]])
+        self.assertEqual([item.title for item in rewritten], ["OpenAI launches GPT-5 API", "OpenAI launches GPT-5 API"])
+        self.assertEqual(report["fallback_count"], 2)
+
+    def test_rewrite_items_defaults_to_large_batches_to_reduce_llm_requests(self):
+        items = [news_item(str(index)) for index in range(61)]
+        batch_sizes = []
+
+        def llm_call(prompt):
+            payload = json.loads(prompt)
+            batch_sizes.append(len(payload["items"]))
+            return json.dumps(
+                {
+                    "rewrites": [
+                        {
+                            "id": entry["id"],
+                            "title": entry["title_raw"],
+                            "summary": entry["summary_raw"],
+                            "flags": [],
+                        }
+                        for entry in payload["items"]
+                    ]
+                }
+            )
+
+        rewrite_items(items, llm_call=llm_call)
+
+        self.assertEqual(batch_sizes, [30, 30, 1])
+
     def test_rewrite_items_does_not_retry_single_items_after_transient_http_error(self):
         items = [news_item("a"), news_item("b")]
         calls = 0

tests/test_stage6_guide.py

@@ -24,7 +24,7 @@ def news_item(item_id, title, section="模型与能力"):
 
 
 class Stage6GuideTests(unittest.TestCase):
-    def test_generate_guide_returns_theme_and_valid_threads(self):
+    def test_generate_guide_returns_intro_theme_threads_and_conclusion(self):
         items = [
             news_item("a", "GPT-5 API 发布"),
             news_item("b", "Miso One 开源语音模型"),
@@ -33,6 +33,7 @@ class Stage6GuideTests(unittest.TestCase):
         def llm_call(prompt):
             return json.dumps(
                 {
+                    "intro": "今天的 AI 行业继续围绕模型能力、Agent 产品和基础设施演进展开。",
                     "theme": "模型能力继续向 API 和实时语音两端推进。",
                     "threads": [
                         {
@@ -48,13 +49,16 @@ class Stage6GuideTests(unittest.TestCase):
                             "kind": "thread",
                         },
                     ],
+                    "conclusion": "总体看，模型能力正在进入更多产品入口，生态竞争也在继续加速。",
                 },
                 ensure_ascii=False,
             )
 
         guide, report = generate_guide(items, llm_call=llm_call)
 
+        self.assertEqual(guide["intro"], "今天的 AI 行业继续围绕模型能力、Agent 产品和基础设施演进展开。")
         self.assertEqual(guide["theme"], "模型能力继续向 API 和实时语音两端推进。")
+        self.assertEqual(guide["conclusion"], "总体看，模型能力正在进入更多产品入口，生态竞争也在继续加速。")
         self.assertEqual(len(guide["threads"]), 1)
         self.assertEqual(guide["threads"][0]["item_ids"], ["a", "b"])
         self.assertEqual(report["dropped_thread_count"], 1)
@@ -67,7 +71,9 @@ class Stage6GuideTests(unittest.TestCase):
 
         guide, report = generate_guide(items, llm_call=llm_call)
 
+        self.assertEqual(guide["intro"], "")
         self.assertEqual(guide["theme"], "")
+        self.assertEqual(guide["conclusion"], "")
         self.assertEqual(guide["threads"], [])
         self.assertTrue(report["fallback_used"])
         self.assertIn("TimeoutError", report["errors"][0])

tests/test_stage7_assemble.py

@@ -23,12 +23,13 @@ def news_item(item_id, title, section):
 
 
 class Stage7AssembleTests(unittest.TestCase):
-    def test_assemble_markdown_renders_sections_and_daily_threads(self):
+    def test_assemble_markdown_renders_intro_sections_daily_threads_and_conclusion(self):
         items = [
             news_item("a", "GPT-5 API 发布", "模型与能力"),
             news_item("b", "Anthropic 提交 IPO 文件", "公司与资本"),
         ]
         guide = {
+            "intro": "今天的 AI 行业继续围绕模型、产品和资本展开。",
             "theme": "> 模型和资本两条线都在推进。[1]",
             "threads": [
                 {
@@ -38,10 +39,12 @@ class Stage7AssembleTests(unittest.TestCase):
                     "kind": "thread",
                 }
             ],
+            "conclusion": "总体看，AI 竞争继续从单点模型能力转向产品、基础设施和资本协同。",
         }
 
         md, report = assemble_markdown(items, guide)
 
+        self.assertTrue(md.startswith("## 引言\n\n> 今天的 AI 行业继续围绕模型、产品和资本展开。"))
         self.assertIn("## 导览", md)
         self.assertIn("> 模型和资本两条线都在推进。", md)
         self.assertIn("## 模型与能力", md)
@@ -49,6 +52,7 @@ class Stage7AssembleTests(unittest.TestCase):
         self.assertIn("**2. Anthropic 提交 IPO 文件**", md)
         self.assertIn("## 今日脉络", md)
         self.assertIn("- **模型能力产品化**", md)
+        self.assertTrue(md.endswith("## 总结\n\n> 总体看，AI 竞争继续从单点模型能力转向产品、基础设施和资本协同。"))
         self.assertNotIn("> >", md)
         self.assertNotIn("[1]", md)
         self.assertEqual(report["item_count"], 2)