Citation

Source Types and Citation Formats

ChromaDB Sources

Documents from internal knowledge base (catalog, faq, troubleshooting collections):

Document structure:

{
    'document': 'product_id: SKU001 | name: TechBook Pro...',
    'collection': 'catalog',  # 'faq' or 'troubleshooting'
    'metadata': {
        'source': 'techmart_catalog.csv',
        'row_index': 0
    }
}

Citation format:

• "Based on our product catalog..."
• "According to our FAQ..."
• "From our troubleshooting guide..."
• Sources: ["techmart_catalog.csv", "techmart_faq.csv"]
• Collections: ["catalog", "faq"]

Web Search Sources

Documents from external web search (Exa API):

Document structure:

{
    'document': '<web page content>',
    'collection': 'web_search',
    'source': 'https://techradar.com/article',  # URL
    'title': 'Best Gaming Laptops 2024',
    'author': 'John Doe',
    'published_date': '2024-03-15'
}

Citation format:

• Markdown hyperlinks: [Title](URL)
• Include publication date when available
• Sources: ["https://techradar.com/article", "https://pcmag.com/review"]
• Collections: ["web_search"]

Examples

Example 1: ChromaDB citations only

# Input documents
documents = [
    {'collection': 'catalog', 'metadata': {'source': 'techmart_catalog.csv'}},
    {'collection': 'faq', 'metadata': {'source': 'techmart_faq.csv'}}
]

# Output
{
    'sources': ['techmart_catalog.csv', 'techmart_faq.csv'],
    'collections_used': ['catalog', 'faq'],
    'source_types': ['chromadb', 'chromadb'],
    'citation_format': 'internal_kb'
}

Example 2: Web search citations only

# Input documents
documents = [
    {
        'collection': 'web_search',
        'source': 'https://techradar.com/best-laptops',
        'title': 'Best Gaming Laptops 2024'
    },
    {
        'collection': 'web_search',
        'source': 'https://pcmag.com/laptop-review',
        'title': 'Gaming Laptop Reviews'
    }
]

# Output
{
    'sources': [
        'https://techradar.com/best-laptops',
        'https://pcmag.com/laptop-review'
    ],
    'collections_used': ['web_search'],
    'source_types': ['web', 'web'],
    'citation_format': 'web_urls',
    'formatted_citations': [
        '[Best Gaming Laptops 2024](https://techradar.com/best-laptops)',
        '[Gaming Laptop Reviews](https://pcmag.com/laptop-review)'
    ]
}

Example 3: Mixed sources (ChromaDB + Web)

# Input documents
documents = [
    {'collection': 'catalog', 'metadata': {'source': 'techmart_catalog.csv'}},
    {
        'collection': 'web_search',
        'source': 'https://example.com/article',
        'title': 'External Review'
    }
]

# Output
{
    'sources': [
        'techmart_catalog.csv',
        'https://example.com/article'
    ],
    'collections_used': ['catalog', 'web_search'],
    'source_types': ['chromadb', 'web'],
    'citation_format': 'mixed',
    'formatted_citations': [
        'Product Catalog',
        '[External Review](https://example.com/article)'
    ]
}

Example 4: Citation validation

# Validate that generated answer only uses information from provided documents
answer = "The TechBook Pro costs $1,499"
documents = [
    {'document': 'product_id: SKU001 | price: 1499.0 | name: TechBook Pro'}
]

validation = validate_citations(answer, documents)
# Returns: {'valid': True, 'grounded': True, 'hallucination_detected': False}

Usage in Generator Agent

The citation skill is used AFTER generation to format and validate sources:

# Step 1: Generate answer (generation_skill)
answer_result = generate_answer(query, documents)

# Step 2: Format citations (citation_skill)
citation_result = format_citations(documents)

# Step 3: Combine results
final_result = {
    'answer': answer_result['answer'],
    'sources': citation_result['sources'],
    'collections_used': citation_result['collections_used'],
    'source_types': citation_result['source_types'],
    'formatted_citations': citation_result.get('formatted_citations', [])
}

Quality Checks

The citation skill performs these quality validations:

1. Source completeness - All documents have valid source information
2. Format consistency - Citations follow correct format for source type
3. URL validation - Web URLs are properly formed
4. Collection verification - Collection names are valid
5. No hallucination - Answer content is grounded in provided documents

Critical Implementation Notes

• Source type detection: Check collection == 'web_search' to identify web sources
• Unique sources: Deduplicate sources list (multiple docs from same CSV/URL)
• Empty documents: Handle gracefully when no documents provided
• Metadata extraction: Safely handle missing metadata fields (author, published_date)
• Markdown formatting: Web citations use [Title](URL) format for proper rendering