Generation

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:

1. With context (documents provided): Synthesizes from multiple documents with source attribution
2. 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