rewriting

edangx100 的头像作者 edangx100

通过上下文感知的改进增强模糊查询以获得更好的检索效果。在没有找到相关文档时进行自适应重试、查询不清晰时或用户提及查询增强或重写时使用。

📖 中文导读:通过上下文感知的改进增强模糊查询以获得更好的检索效果。在没有找到相关文档时进行自适应重试、查询不清晰时或用户提及查询增强或重写时使用。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: rewriting description: Enhances vague queries for better retrieval with context-aware improvements. Use during adaptive retry when no relevant documents found, when queries are unclear, or when the user mentions query enhancement or rewriting.

Query Rewriting

Instructions

Rewrite queries using rewrite_query() in components/rewriter.py. Typically triggered when len(relevant_docs) == 0 during adaptive retry.

Default usage:

# During retry after no relevant docs found
previous_context = f"Previous query '{current_query}' found {len(docs)} documents but none were relevant."

rewrite_result = rewrite_query(current_query, previous_context=previous_context)
current_query = rewrite_result['rewritten_query']

Rewriting strategies:

  • Expand abbreviations: "fast computer" → "high performance laptop processor"
  • Clarify vague terms: "won't work" → "troubleshooting device not powering on"
  • Incorporate context: "What about warranty" + product context → "warranty coverage for [product]"
  • Add synonyms: "cheap" → "affordable budget inexpensive"

Returns:

  • rewritten_query: Enhanced query string
  • reasoning: Explanation of changes

Critical for adaptive retry: Call only when len(relevant_docs) == 0 and num_retries < RETRY_LIMIT. Pass context from previous failed attempt.

Implementation: components/rewriter.py, uses REWRITE_MODEL from config.py (default: Haiku 4.5), temperature 0.3.

Examples

Example 1: Basic rewriting (no context)

# Input
result = rewrite_query("fast computer")

# Output
{
    "rewritten_query": "high performance laptop with fast processor and SSD storage",
    "reasoning": "Expanded 'fast' to specific performance characteristics"
}

Example 2: Context-aware rewriting

# Input
result = rewrite_query(
    "What about warranty",
    previous_context="User interested in ZenithBook 13 Evo laptop"
)

# Output
{
    "rewritten_query": "What is the warranty coverage and terms for ZenithBook 13 Evo laptop",
    "reasoning": "Incorporated product context and made question explicit"
}

Example 3: Adaptive retry integration

# After grading finds no relevant docs
if len(relevant_documents) == 0 and num_retries < RETRY_LIMIT:
    # Build context
    previous_context = f"Previous query '{current_query}' found {len(retrieved_documents)} documents but none were relevant."

    # Rewrite
    rewrite_result = rewrite_query(current_query, previous_context)
    current_query = rewrite_result['rewritten_query']

    print(f"Retry {num_retries + 1}: {current_query}")
    num_retries += 1
    # Loop back to retrieval

Example 4: Vague troubleshooting

# Input
result = rewrite_query("won't work")

# Output
{
    "rewritten_query": "troubleshooting device not functioning properly or not turning on",
    "reasoning": "Clarified vague complaint into specific troubleshooting query"
}