From 5ce60cbd957fc5bc97946c365591b1146e7aa8aa Mon Sep 17 00:00:00 2001 From: liuwei Date: Wed, 15 Apr 2026 10:37:27 +0800 Subject: [PATCH] =?UTF-8?q?=E6=88=91=E5=B7=B2=E7=BB=8F=E6=8A=8A=E7=BE=A4?= =?UTF-8?q?=E6=80=BB=E7=BB=93=E8=BF=99=E5=9D=97=E4=BB=8E=E2=80=9C=E5=89=8D?= =?UTF-8?q?=20120=20=E5=AD=97=E7=A1=AC=E7=A0=8D=E2=80=9D=E6=94=B9=E6=88=90?= =?UTF-8?q?=E4=BA=86=E2=80=9C=E5=85=B3=E9=94=AE=E5=8F=A5=E6=8F=90=E7=82=BC?= =?UTF-8?q?=20+=20=E6=9B=B4=E9=AB=98=E4=B8=8A=E9=99=90=E2=80=9D=E3=80=82?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 已改文件： context_builder.py 具体优化： 原来： group_memory_summary 超过 120 字就直接 summary[:117] + "..."，很容易把后面的关键结论截掉。 现在： 新增 _compact_group_summary(...)。 先按句拆分，再给句子打分，优先保留包含“结论/风险/报错/配置/策略/优化”等关键词的句子。 同时保留首句和末句，避免上下文断裂。 最终摘要上限提高到 420 字、最多 6 句，再做兜底截断。 prompt 字段名也改成 群长期摘要关键句，让模型更明确这是提炼后的重点。 --- .../context/context_builder.py | 50 +++++++++++++++++-- 1 file changed, 46 insertions(+), 4 deletions(-) diff --git a/plugins/ai_auto_response/context/context_builder.py b/plugins/ai_auto_response/context/context_builder.py index 6a3c293..3eeb484 100644 --- a/plugins/ai_auto_response/context/context_builder.py +++ b/plugins/ai_auto_response/context/context_builder.py @@ -276,9 +276,7 @@ class ContextBuilder: return "当前群没有特殊知识域限制。" focus = ", ".join(group_profile.get("knowledge_focus", [])[:6]) boundaries = ", ".join(group_profile.get("topic_boundaries", [])[:6]) - summary = str(group_profile.get("group_memory_summary", "") or "").replace("\n", " ").strip() - if len(summary) > 120: - summary = summary[:117] + "..." + summary = ContextBuilder._compact_group_summary(str(group_profile.get("group_memory_summary", "") or "")) lines = [ f"群模式：{group_profile.get('mode', 'social')}", f"知识域偏向：{group_profile.get('knowledge_domain', 'general')}（仅作理解倾向，不是每次都要显式提到）", @@ -291,7 +289,7 @@ class ContextBuilder: f"表达松弛度：{group_profile.get('expressiveness_style', '克制')}", f"称呼强度：{group_profile.get('address_style', '低频称呼，默认直接接话')}", f"可能相关的话题背景：{focus}" if focus else "", - f"群长期摘要：{summary}" if summary else "", + f"群长期摘要关键句：{summary}" if summary else "", f"历史推断社交风格：{ContextBuilder._build_style_summary(group_profile.get('group_memory_style', {}))}" if group_profile.get("group_memory_style") else "", @@ -313,6 +311,50 @@ class ContextBuilder: ] ).strip(" /") + @staticmethod + def _compact_group_summary(summary_text: str, max_chars: int = 420, max_sentences: int = 6) -> str: + text = str(summary_text or "").strip() + if not text: + return "" + text = re.sub(r"\s+", " ", text.replace("\n", " ").replace("\r", " ")).strip() + if len(text) <= max_chars: + return text + + # 长摘要按句提炼，优先保留“结论/风险/动作/配置”等关键词句，避免简单截断丢重点。 + sentences = [part.strip(" ，,；;。.!?！？:：") for part in re.split(r"[。！？!?；;\n]+", text) if part.strip()] + if not sentences: + return text[: max_chars - 3] + "..." + + key_patterns = [ + r"结论|核心|重点|关键|建议|方案|步骤|原因|影响|风险|注意|问题|异常|报错|故障|超时|阻塞", + r"配置|参数|阈值|策略|限制|回退|优化|修复|排查|上线|回滚|依赖|版本|兼容", + ] + scored: List[tuple[int, int, str]] = [] + for idx, sentence in enumerate(sentences): + score = 0 + for pattern in key_patterns: + if re.search(pattern, sentence, flags=re.IGNORECASE): + score += 3 + if re.search(r"\d", sentence): + score += 1 + if 8 <= len(sentence) <= 80: + score += 1 + if idx == 0 or idx == len(sentences) - 1: + score += 1 + scored.append((score, idx, sentence)) + + chosen_indexes = {0, len(sentences) - 1} + for _, idx, _ in sorted(scored, key=lambda x: (-x[0], x[1])): + chosen_indexes.add(idx) + if len(chosen_indexes) >= max_sentences: + break + + chosen = [sentences[idx] for idx in sorted(chosen_indexes)] + merged = "；".join([item for item in chosen if item]).strip("；") + if len(merged) <= max_chars: + return merged + return merged[: max_chars - 3].rstrip(" ，,；;。.!?！？:：") + "..." + @staticmethod def _build_quote_prompt(quote_context: Dict) -> str: if not quote_context: