Python SDK Guide

The Python SDK is designed for seamless integration into backend services, data pipelines, and AI agent frameworks. It provides programmatic access to the full CiteKit lifecycle.

Initialization

The CiteKitClient is the main entry point. It manages the mapper (Gemini by default or a custom MapperProvider) and the local file system (for storage and resolution).

from citekit import CiteKitClient

# Default Initialization
# Uses '.resource_maps' and '.citekit_output' in current dir
client = CiteKitClient()

# Custom Configuration
client = CiteKitClient(
    base_dir="./my_library",     # Root for relative paths
    storage_dir="./maps",        # Where JSON maps are stored
    output_dir="./media",        # Where resolved files go
    concurrency_limit=10,        # Max API calls
    api_key="AIzaSy..."          # Explicit API Key (default: env.GEMINI_API_KEY)
)

1. Ingestion (Mapping)

The ingest method analyzes a file and returns a ResourceMap object.

async def analyze_file():
    # Returns a ResourceMap Pydantic model
    resource_map = await client.ingest(
        resource_path="lectures/cs101.mp4",
        resource_type="video",  # Optional: 'video', 'document', 'audio', 'image', 'text'
        resource_id="cs101_lec1" # Optional: Custom ID (default: filename)
    )
    
    print(f"Title: {resource_map.title}")
    print(f"Total Nodes: {len(resource_map.nodes)}")

Error Handling:

• FileNotFoundError: If the input file doesn't exist.
• ValueError: If the file type is unsupported or prompt fails.
• Exception: API connection errors (CiteKit handles retries automatically).

2. Inspection (Planning)

Once mapped, you can retrieve and inspect the structure without hitting the API. This is heavily used by Agents to "see" what's available.

# Retrieve an existing map by ID
resource_map = client.get_map("cs101_lec1")

# Iterate through top-level nodes
for node in resource_map.nodes:
    print(f"[{node.id}] {node.title} ({node.type})")
    
    # Check if it has children (nested structure)
    if node.children:
        for child in node.children:
             print(f"  - {child.title}")

Accessing Metadata

Nodes contain rich metadata about their location in the source file:

node = resource_map.nodes[0]

# For Video/Audio
start_time = node.location.start  # Seconds
end_time = node.location.end      # Seconds

# For Documents (PDF)
pages = node.location.pages

3. Resolution (Extraction)

Resolution converts a logical node_id into a physical file or actionable metadata.

Physical Resolution

Creates a new file on disk containing only the relevant content.

evidence = client.resolve(
    resource_id="cs101_lec1", 
    node_id="intro_chapter"
)

# The path to the newly created clip/slice
print(f"Generated: {evidence.output_path}")
# e.g., ./media/cs101_lec1_intro_chapter.mp4

Virtual Resolution

Returns valid addresses and metadata without running extraction tools (FFmpeg/PDF).

evidence = client.resolve(
    resource_id="cs101_lec1", 
    node_id="intro_chapter",
    virtual=True
)

print(evidence.output_path) # None
print(evidence.address)     # video://cs101_lec1#t=645-1350