This Claude skill serves as the final synthesis engine for Retrieval-Augmented Generation (RAG) pipelines. It intelligently aggregates information from multiple retrieved documents to produce coherent, cited answers. By prioritizing the original user query and providing structured metadata—including source attribution and fallback mechanisms—it ensures high-quality, grounded responses even when information is scarce.

When should I use generation?

generation is useful in the following scenarios: • Customer Support Automation: Powering help centers by synthesizing answers from product manuals and FAQs with direct citations to source documents. • Internal Knowledge Retrieval: Enabling employees to query complex corporate wikis or HR policies and receive summarized, factual answers with clear provenance. • Technical Documentation Synthesis: Helping developers navigate extensive API documentation by combining relevant snippets into a single, actionable response. • Research Assistance: Aggregating findings from multiple academic papers or market reports into a cohesive summary that attributes every claim to its specific source.

name	generation
description	Generates answers by synthesizing information from retrieved documents with citations. Use as final RAG pipeline step, when creating answers from multiple sources, or when the user mentions answer generation or synthesis.

Answer Generation

Instructions

Generate answers using functions in components/generator.py. Synthesizes information from graded-relevant documents or provides fallback when no context available.

Default workflow:

# Use original query (not rewritten) for generation
result = generate_answer_with_metadata(original_query, relevant_documents)
answer = result['answer']

Key functions:

# Standard generation with metadata (preferred)
result = generate_answer_with_metadata(query, documents, include_sources=True)
# Returns: answer, num_documents_used, has_context, sources, collections_used

# Basic generation
result = generate_answer(query, documents)
# Returns: answer, num_documents_used, has_context, model_used

Generation modes:

With context (documents provided): Synthesizes from multiple documents with source attribution
Without context (no documents or empty list): Acknowledges lack of information, provides helpful fallback

Critical: Always use original_query for generation, NOT current_query (which may be rewritten). This ensures the answer addresses what the user actually asked.

Implementation: components/generator.py, uses GENERATION_MODEL from config.py (default: Haiku 4.5), temperature 0.3, max tokens 2000.

Examples

Example 1: Standard generation with context

# Input
result = generate_answer_with_metadata(
    "What gaming laptops do you have?",
    relevant_documents  # 3 documents from catalog
)

# Output
{
    "answer": "We have several gaming laptops including TechBook Pro 15 with RTX 4060...",
    "num_documents_used": 3,
    "has_context": True,
    "sources": ["techmart_catalog.csv"],
    "collections_used": ["catalog"]
}

Example 2: Fallback (no documents)

# Input
result = generate_answer("obscure query", [])

# Output
{
    "answer": "I don't have relevant information about this in our knowledge base...",
    "num_documents_used": 0,
    "has_context": False,
    "sources": [],
    "collections_used": []
}

Example 3: Pipeline integration

# After retrieval, grading, retries
if relevant_documents:
    # Generate with context
    generation_result = generate_answer_with_metadata(
        original_query,  # NOT current_query
        relevant_documents
    )
    print(f"✓ Answer from {generation_result['num_documents_used']} docs")
else:
    # Fallback after exhausting retries
    generation_result = generate_answer(original_query, [])
    print("⚠ Fallback answer (no relevant docs)")

print(generation_result['answer'])

Example 4: Using original vs rewritten query

# Setup
original_query = "gaming laptops"
current_query = "high performance gaming laptop computers RTX graphics"  # Rewritten

# After retrieval with current_query
relevant_docs = filter_relevant_documents(graded_docs)

# Generate with ORIGINAL query
result = generate_answer(original_query, relevant_docs)  # ✓ Correct
# NOT: generate_answer(current_query, relevant_docs)     # ✗ Wrong

generation

When & Why to Use This Skill

Use Cases