graphics-llm / src /datawrapper_client.py
Tom
Add Datawrapper chart generation mode with clean iframe display
7114af0
raw
history blame
11.3 kB
"""
Datawrapper Chart Generation Client
Integrates RAG pipeline with Datawrapper API for intelligent chart creation.
"""
import json
import os
from typing import Optional, Tuple
import pandas as pd
from .prompts import (
CHART_SELECTION_SYSTEM_PROMPT,
get_chart_selection_prompt,
get_chart_styling_prompt
)
from .llm_client import create_llm_client
from .rag_pipeline import GraphicsDesignPipeline
# Import Datawrapper MCP handlers directly
from datawrapper_mcp.handlers.create import create_chart as mcp_create_chart
from datawrapper_mcp.handlers.publish import publish_chart as mcp_publish_chart
from datawrapper_mcp.handlers.retrieve import get_chart_info as mcp_get_chart_info
def get_data_summary(df: pd.DataFrame) -> str:
"""
Generate a summary of the DataFrame structure and content.
Args:
df: Input DataFrame
Returns:
String summary of data characteristics
"""
summary_parts = []
# Basic info
summary_parts.append(f"Rows: {len(df)}, Columns: {len(df.columns)}")
summary_parts.append(f"Column names: {', '.join(df.columns.tolist())}")
# Column types
numeric_cols = df.select_dtypes(include=['number']).columns.tolist()
text_cols = df.select_dtypes(include=['object']).columns.tolist()
date_cols = df.select_dtypes(include=['datetime']).columns.tolist()
if numeric_cols:
summary_parts.append(f"Numeric columns: {', '.join(numeric_cols)}")
if text_cols:
summary_parts.append(f"Text columns: {', '.join(text_cols)}")
if date_cols:
summary_parts.append(f"Date columns: {', '.join(date_cols)}")
# Data preview (first 3 rows)
summary_parts.append(f"\nData preview:\n{df.head(3).to_string()}")
return "\n".join(summary_parts)
def analyze_csv_for_chart_type(
df: pd.DataFrame,
user_prompt: str,
rag_pipeline: GraphicsDesignPipeline
) -> Tuple[str, str]:
"""
Use RAG and LLM to determine the best chart type for the data.
Args:
df: Input DataFrame
user_prompt: User's description of what they want to visualize
rag_pipeline: RAG pipeline for retrieving best practices
Returns:
Tuple of (chart_type, reasoning)
"""
# Get data summary
data_summary = get_data_summary(df)
# Query RAG for chart selection best practices
rag_query = f"chart type selection for {user_prompt}"
relevant_docs = rag_pipeline.retrieve_documents(rag_query, k=3)
rag_context = rag_pipeline.vectorstore.format_documents_for_context(relevant_docs)
# Generate chart type recommendation using LLM
chart_prompt = get_chart_selection_prompt()
full_prompt = chart_prompt.format(
user_prompt=user_prompt,
data_summary=data_summary,
rag_context=rag_context
)
llm_client = create_llm_client(
model=os.getenv("LLM_MODEL", "meta-llama/Llama-3.1-8B-Instruct"),
temperature=0.3, # Lower temperature for more deterministic chart selection
max_tokens=500
)
response = llm_client.generate(
prompt=full_prompt,
system_prompt=CHART_SELECTION_SYSTEM_PROMPT
)
# Parse JSON response
try:
# Extract JSON from response (handle markdown code blocks)
response_clean = response.strip()
if "```json" in response_clean:
response_clean = response_clean.split("```json")[1].split("```")[0].strip()
elif "```" in response_clean:
response_clean = response_clean.split("```")[1].split("```")[0].strip()
result = json.loads(response_clean)
chart_type = result.get("chart_type", "line")
reasoning = result.get("reasoning", "")
# Validate chart type
valid_types = ["bar", "line", "area", "scatter", "column", "stacked_bar", "arrow", "multiple_column"]
if chart_type not in valid_types:
chart_type = "line" # Default fallback
return chart_type, reasoning
except Exception as e:
print(f"Error parsing chart type response: {e}")
print(f"Response was: {response}")
# Default to line chart
return "line", "Using default line chart due to parsing error"
def generate_chart_config(
chart_type: str,
df: pd.DataFrame,
user_prompt: str,
rag_pipeline: GraphicsDesignPipeline
) -> dict:
"""
Generate Datawrapper chart configuration using RAG and LLM.
Args:
chart_type: Type of chart to create
df: Input DataFrame
user_prompt: User's visualization request
rag_pipeline: RAG pipeline for retrieving design best practices
Returns:
Dictionary with chart configuration
"""
# Get data summary
data_summary = get_data_summary(df)
# Query RAG for styling and design best practices
rag_query = f"chart design best practices colors accessibility {chart_type}"
relevant_docs = rag_pipeline.retrieve_documents(rag_query, k=3)
rag_context = rag_pipeline.vectorstore.format_documents_for_context(relevant_docs)
# Generate chart configuration using LLM
styling_prompt = get_chart_styling_prompt()
full_prompt = styling_prompt.format(
chart_type=chart_type,
user_prompt=user_prompt,
data_summary=data_summary,
rag_context=rag_context
)
llm_client = create_llm_client(
model=os.getenv("LLM_MODEL", "meta-llama/Llama-3.1-8B-Instruct"),
temperature=0.5,
max_tokens=800
)
response = llm_client.generate(
prompt=full_prompt,
system_prompt="You are a data visualization expert. Generate valid JSON configuration for Datawrapper charts."
)
# Parse JSON response
try:
# Extract JSON from response
response_clean = response.strip()
if "```json" in response_clean:
response_clean = response_clean.split("```json")[1].split("```")[0].strip()
elif "```" in response_clean:
response_clean = response_clean.split("```")[1].split("```")[0].strip()
config = json.loads(response_clean)
# Ensure basic required fields
if "title" not in config:
config["title"] = user_prompt[:100] # Use prompt as fallback title
return config
except Exception as e:
print(f"Error parsing chart config: {e}")
print(f"Response was: {response}")
# Return minimal config
return {
"title": user_prompt[:100] if user_prompt else "Data Visualization",
"source_name": "User Data"
}
async def create_and_publish_chart(
df: pd.DataFrame,
user_prompt: str,
rag_pipeline: GraphicsDesignPipeline,
api_token: Optional[str] = None
) -> dict:
"""
Complete workflow: analyze data, select chart type, create and publish chart.
Args:
df: Input DataFrame
user_prompt: User's visualization request
rag_pipeline: RAG pipeline instance
api_token: Datawrapper API token (defaults to env var)
Returns:
Dictionary with chart info including iframe URL
"""
if api_token is None:
api_token = os.getenv("DATAWRAPPER_ACCESS_TOKEN")
if not api_token:
raise ValueError("DATAWRAPPER_ACCESS_TOKEN not found in environment")
try:
# Step 1: Analyze data and select chart type
chart_type, reasoning = analyze_csv_for_chart_type(df, user_prompt, rag_pipeline)
# Step 2: Generate chart configuration
chart_config = generate_chart_config(chart_type, df, user_prompt, rag_pipeline)
# Step 3: Convert DataFrame to list of dicts for Datawrapper
data_list = df.to_dict('records')
# Step 4: Create chart using MCP handler
create_args = {
"data": data_list,
"chart_type": chart_type,
"chart_config": chart_config
}
create_result = await mcp_create_chart(create_args)
if not create_result or len(create_result) == 0:
raise ValueError("Empty response from chart creation")
result_text = create_result[0].text
if not result_text or result_text.strip() == "":
raise ValueError("Empty text in chart creation response")
result_data = json.loads(result_text)
chart_id = result_data.get("chart_id")
if not chart_id:
raise ValueError(f"Failed to get chart_id from creation response. Response was: {result_data}")
# Step 5: Try to publish chart using MCP handler
publish_success = False
publish_message = ""
try:
publish_args = {"chart_id": chart_id}
publish_result = await mcp_publish_chart(publish_args)
publish_text = publish_result[0].text
publish_data = json.loads(publish_text)
publish_success = True
publish_message = publish_data.get("message", "Published successfully")
except Exception as publish_error:
publish_message = f"Publish failed: {str(publish_error)}"
# Step 6: Get full chart info using MCP handler
chart_info_args = {"chart_id": chart_id}
chart_info_result = await mcp_get_chart_info(chart_info_args)
chart_info_text = chart_info_result[0].text
chart_info = json.loads(chart_info_text)
# Return complete info
return {
"success": True,
"chart_id": chart_id,
"chart_type": chart_type,
"reasoning": reasoning,
"public_url": chart_info.get("public_url"),
"edit_url": chart_info.get("edit_url"),
"published": publish_success,
"publish_message": publish_message,
"title": chart_config.get("title", "Chart")
}
except json.JSONDecodeError as e:
error_msg = f"JSON parsing error: {str(e)}"
print(f"Error in chart creation: {error_msg}")
print(f"Failed to parse: {result_text if 'result_text' in locals() else 'N/A'}")
return {
"success": False,
"error": error_msg,
"chart_type": chart_type if 'chart_type' in locals() else None,
"public_url": None
}
except Exception as e:
error_msg = f"{type(e).__name__}: {str(e)}"
print(f"Error in chart creation: {error_msg}")
import traceback
traceback.print_exc()
return {
"success": False,
"error": error_msg,
"chart_type": chart_type if 'chart_type' in locals() else None,
"public_url": None
}
def get_iframe_html(chart_url: str, height: int = 600) -> str:
"""
Generate iframe HTML for embedding a Datawrapper chart.
Args:
chart_url: Public URL of the chart
height: Height of iframe in pixels
Returns:
HTML string with iframe
"""
if not chart_url:
return "<div style='padding: 50px; text-align: center;'>No chart available</div>"
return f"""
<div style="width: 100%; height: {height}px;">
<iframe
src="{chart_url}"
style="width: 100%; height: 100%; border: none;"
frameborder="0"
scrolling="no"
aria-label="Chart">
</iframe>
</div>
"""