Skip to main content

Quick Insights

Quick Insights are automated statistical profiles and conversational analytics interfaces generated immediately upon data ingestion. They are designed for quick, lightweight exploration of a single file or data source without requiring coding skills.

Target Persona & Use Case​

  • Who it is for: Basic Users and business stakeholders (e.g., Operations Managers, Product Managers, or business analysts) who want immediate answers from a dataset.
  • The Goal: To get a high-level view and ask basic questions of a newly uploaded dataset without writing SQL, Python, or configuring complex notebooks.
  • Interaction: Simple plain-English questions regarding a single data source, coupled with automated summary statistics and quality diagnostics.

Technical Flow​

When a new dataset (CSV, Parquet, JSON) is uploaded to the warehouse, Ravioli executes a pre-configured profiling pipeline:

1. Data Cleaning and Type Casting​

Before profiling, the dataset is cleaned to prevent calculation noise:

  • ID Columns: Standard ID patterns (e.g., _id, postcode, phone) are detected and cast to categories or strings, preventing the calculator from computing meaningless averages.
  • Low-Cardinality Categories: Text columns with fewer than 30 unique values are optimized to category types to save memory.

2. Statistical Profiling​

Ravioli runs the ydata-profiling engine in minimal mode to analyze:

  • Row and column dimensions.
  • Missing values and zero counts.
  • High-level skewness and correlations.

3. LLM Synthesis & Conversational Interface​

The statistical profile is sent to the LLM agent, which maps the properties to quick_insight_template.md to produce an initial report with:

  • Key Findings: Statistics and percentages highlighted in backticks.
  • Governance Assumptions: Grounding rules for data validity.
  • Data Constraints: Skewness warnings or empty columns.

Based on this report, Ravioli automatically suggests 3 follow-up prompts to launch conversational query sessions, enabling basic users to probe the data dynamically.