Improve LLM rewrite classification pipeline

ai_daily_report/rewrite.py

@@ -4,6 +4,7 @@ import json
 from typing import Any, Callable
 from urllib.error import HTTPError
 
+from .classify import SECTION_ORDER
 from .llm import parse_json_object
 from .models import NewsItem
 
@@ -18,9 +19,21 @@ def _chunks(items: list[NewsItem], size: int) -> list[list[NewsItem]]:
 def _build_prompt(batch: list[NewsItem]) -> str:
     payload = {
         "task": (
-            "Rewrite AI news titles and summaries into concise Chinese. Preserve brand/model/API names "
-            "such as GPT-5, Codex, Gemini, Claude, API, MCP. Do not add facts."
+            "For each AI news item, translate when needed, rewrite the title and summary into concise Chinese, "
+            "and classify it into exactly one allowed section. Preserve brand/model/API names such as GPT-5, "
+            "Codex, Gemini, Claude, API, MCP. Do not add facts."
         ),
+        "allowed_sections": SECTION_ORDER,
+        "section_guidance": {
+            "模型与能力": "model releases, capability upgrades, modalities, context windows, inference, benchmarks tied to model ability",
+            "产品与应用": "end-user products, apps, agents, workflows, product launches, practical business or consumer use cases",
+            "开发与基础设施": "developer tools, APIs, SDKs, MCP, frameworks, deployment, chips, cloud, infra, open source engineering",
+            "公司与资本": "company strategy, financing, IPO, acquisitions, partnerships, revenue, business competition",
+            "政策与安全": "policy, regulation, safety, privacy, copyright, misuse, security incidents, governance",
+            "论文与研究": "papers, academic research, arXiv, methods, experiments, datasets, evaluations",
+            "观点与教程": "opinions, analysis, explainers, tutorials, guides, practices",
+            "人物与动态": "people-focused interviews, speeches, career moves, public appearances",
+        },
         "items": [
             {
                 "id": item.id,
@@ -28,6 +41,7 @@ def _build_prompt(batch: list[NewsItem]) -> str:
                 "summary_raw": item.summary_raw,
                 "source": item.source_label,
                 "language_hint": item.language_hint,
+                "source_section_hint": item.section_hint,
             }
             for item in batch
         ],
@@ -37,6 +51,8 @@ def _build_prompt(batch: list[NewsItem]) -> str:
                     "id": "item id",
                     "title": "display title",
                     "summary": "display summary",
+                    "section": "one allowed section",
+                    "confidence": 0.0,
                     "flags": [],
                 }
             ]
@@ -58,13 +74,14 @@ def _is_transient_llm_error(exc: Exception) -> bool:
     return False
 
 
-def _apply_rewrite_batch(batch: list[NewsItem], llm_call: RewriteLlmCall) -> int:
+def _apply_rewrite_batch(batch: list[NewsItem], llm_call: RewriteLlmCall) -> tuple[int, int]:
     obj = parse_json_object(llm_call(_build_prompt(batch)))
     rewrites = obj.get("rewrites", [])
     if not isinstance(rewrites, list):
         raise ValueError("rewrites is not a list")
     by_id = {item.id: item for item in batch}
     seen_ids: set[str] = set()
+    section_count = 0
     for entry in rewrites:
         item_id = entry.get("id")
         title = str(entry.get("title") or "").strip()
@@ -72,8 +89,12 @@ def _apply_rewrite_batch(batch: list[NewsItem], llm_call: RewriteLlmCall) -> int
         if item_id in by_id and title and summary:
             by_id[item_id].title = title
             by_id[item_id].summary = summary
+            section = str(entry.get("section") or "").strip()
+            if section in SECTION_ORDER:
+                by_id[item_id].section = section
+                section_count += 1
             seen_ids.add(item_id)
-    return len(seen_ids)
+    return len(seen_ids), section_count
 
 
 def rewrite_items(
@@ -85,14 +106,16 @@ def rewrite_items(
     retry_single_items: bool = False,
 ) -> tuple[list[NewsItem], dict[str, Any]]:
     rewritten_count = 0
+    llm_section_count = 0
     fallback_count = 0
     missing_rewrite_count = 0
     errors: list[str] = []
 
     for batch in _chunks(items, max(1, batch_size)):
         try:
-            batch_rewritten_count = _apply_rewrite_batch(batch, llm_call)
+            batch_rewritten_count, batch_section_count = _apply_rewrite_batch(batch, llm_call)
             rewritten_count += batch_rewritten_count
+            llm_section_count += batch_section_count
             for item in batch:
                 if item.title is None or item.summary is None:
                     errors.append(f"missing_rewrite_for_item: {item.id}")
@@ -113,7 +136,9 @@ def rewrite_items(
                 continue
             for item in batch:
                 try:
-                    rewritten_count += _apply_rewrite_batch([item], llm_call)
+                    item_rewritten_count, item_section_count = _apply_rewrite_batch([item], llm_call)
+                    rewritten_count += item_rewritten_count
+                    llm_section_count += item_section_count
                 except Exception as item_exc:
                     errors.append(f"item:{item.id}:{type(item_exc).__name__}: {item_exc}")
                     _fallback(item)
@@ -127,6 +152,7 @@ def rewrite_items(
     report = {
         "input_count": len(items),
         "rewritten_count": rewritten_count,
+        "llm_section_count": llm_section_count,
         "fallback_count": fallback_count,
         "missing_rewrite_count": missing_rewrite_count,
         "fallback_ratio": round(fallback_ratio, 4),