knowledge-card-retrieval

Installed Root

Resolve the installed reasflow-dev skills root before running packaged scripts:

REASFLOW_SKILLS_ROOT="${REASFLOW_SKILLS_ROOT:-}"
if [ -z "$REASFLOW_SKILLS_ROOT" ]; then
  if [ -d ./.agents/skills ]; then
    REASFLOW_SKILLS_ROOT="$(pwd)/.agents/skills"
  elif [ -d "$HOME/.agents/skills" ]; then
    REASFLOW_SKILLS_ROOT="$HOME/.agents/skills"
  else
    echo "reasflow shared skills not found in ./.agents/skills or $HOME/.agents/skills" >&2
    exit 1
  fi
fi

REASFLOW_PRIVATE_SKILLS_ROOT="${REASFLOW_PRIVATE_SKILLS_ROOT:-}"
if [ -z "$REASFLOW_PRIVATE_SKILLS_ROOT" ]; then
  if [ -d ./.codex/reasflow-skills ]; then
    REASFLOW_PRIVATE_SKILLS_ROOT="$(pwd)/.codex/reasflow-skills"
  elif [ -d "$HOME/.codex/reasflow-skills" ]; then
    REASFLOW_PRIVATE_SKILLS_ROOT="$HOME/.codex/reasflow-skills"
  else
    echo "reasflow private skills not found in ./.codex/reasflow-skills or $HOME/.codex/reasflow-skills" >&2
    exit 1
  fi
fi

知识卡片检索

概述

你是知识库管理员，负责从领域知识卡片目录中检索最相关的数学技术，以支持证明任务。 你必须使用分层检索策略，确保不遗漏关键技术，从高层框架逐步深入到低层优化技巧。

SKILL_ROOT="$REASFLOW_SKILLS_ROOT/knowledge-card-retrieval"

内置卡片入口："$SKILL_ROOT/assets/knowledge-cards/catalog.json"

检索命令

python "$SKILL_ROOT/scripts/search-cards.py" \
  --catalog "$SKILL_ROOT/assets/knowledge-cards/catalog.json" \
  --query "<查询词>" \
  --top-k 5

# 按 ID 读取卡片全文：
python "$SKILL_ROOT/scripts/read-card.py" --id <CARD_ID>
# 或直接读取路径：
cat "<search 输出中显示的绝对路径>"

---

5层检索协议（The 5-Layer Protocol）

第1层：核心机制搜索（"骨架"）

• 目标：找到主要证明框架和算法特有技术
• 行动：提取核心算法名称和关键机制（如 "FedAvg"、"Momentum"、"Subspace"）
• 查询模板："<算法名> convergence <机制1> <机制2>"
• 示例："FedAvg convergence random subspace momentum"

第2层：结构性质搜索（"特征"）

• 目标：识别结构属性（分布式、并行、动量）和性能保证（线性加速）
• 行动：
  • 分布式/并行：搜索 "Distributed variance reduction", "Linear speedup"
  • 动量/加速：搜索 "Momentum accumulation", "Acceleration technique"
  • 压缩/量化：搜索 "Error feedback", "Compression variance"
• 查询模板："<属性/结构> <关键词> technique"