Skip to main content

AI Analyst: Kowalski

Kowalski the AI Analyst

Kowalski is Ravioli's primary AI Data Analyst agent. Powered by LangChain and implemented in Kowalski.py, Kowalski executes analytical workflows using models configured via the LLM Providers integrations page.

Kowalski contributes directly to three core activities:

  • Analyses: Autonomously drafts execution plans, writes optimized DuckDB SQL queries, self-corrects runtime errors, and synthesizes visual charts.
  • Insights: Summarizes query results into clinical, high-impact bulleted signals ready for organization-wide publication.
  • Data Ingestion: Automatically generates descriptions and profiles structures of newly registered data sources.

Persona & Demeanor​

Kowalski is defined by a strict set of behavioral rules that ensure clinical reliability and an objective presentation of analytical outcomes:

  • Methodical & Precise: Every response is focused on empirical evidence and statistical certainty. Emotional, speculative, or overly descriptive language is avoided in favor of quantifiable metrics.
  • Clinical Tone: Summaries are structured to provide max 4–5 sentences of direct, high-impact findings (Bottom Line Up Front).
  • Analytical Signals: Kowalski punctuates status updates and milestones with Polish analytical confirmations (e.g., "Tak.", "Analiza zakończona.", "Zrozumiałem.", "Gotowe.", "Potwierdzam.") to sign off on specific analytical checkpoints.

Core Capabilities & Skills​

Kowalski's capabilities are divided into five specialized pillars:

1. SQL Synthesis & Optimization​

  • DuckDB Dialect: Fully proficient in high-performance DuckDB SQL, capitalizing on window functions, CTEs, and specific array/struct operations.
  • Schema Mapping: Rapidly interprets table schemas to establish join logic, identify keys, and isolate correct filters.
  • Query Hardening: Synthesizes and self-corrects code before execution to ensure runtime stability.

2. Statistical Reasoning​

  • Trend Analysis: Uncovers patterns, anomalies, and multi-dimensional correlations within raw datasets.
  • Aggregations: Computes statistical summaries (variance, standard deviations, means, and percentiles) to represent dataset distribution profiles accurately.

3. Advanced Data Visualization (VizOps)​

  • Heuristic Selection: Automatically determines the best visual display (Bar, Line, Pie, Scatter, etc.) based on data cardinality and type.
  • Color Theory & High Contrast: Selects accessible, harmonious palettes suited for immediate frontend display.
  • Chart Payloads: Outputs standardized JSON schemas optimized for rendering directly in the frontend via Chart.js.

4. Decision Intelligence​

  • Intent Parsing: Determines whether a question requires a basic textual summary or a multi-step analytical investigation.
  • Resource Optimization: Strategically coordinates and swaps specialized models (e.g., leveraging translation/SQL models vs. general reasoning models) to minimize latency and RAM overhead.

5. Clinical Reporting​

  • Insight Distillation: Distills deep query results into precise, bulleted signals for business consumption.
  • Audit Trails: Generates transparent, step-by-step logs of reasoning steps and query executions.

Specialized Tools​

Kowalski interacts with your data warehouse via a structured set of analytical tools:

ToolPurpose
duckdbThe main engine used to execute raw SQL queries against your local OLAP database.
ingest_fileAutomates the staging and ingestion of external files (CSV, Parquet, JSON, XLSX) into DuckDB.
inspect_tableInvestigates column names, types, and sample data to understand table schemas before writing queries.

Reusable Workflows​

Kowalski standardizes complex analytical routines into three key workflows:

  1. data_ingestion: A methodical pipeline for registering and validating new datasets, checking constraints, and producing schemas.
  2. simple_data_analysis: Quick, conversational querying to extract singular metrics or answer targeted data questions.
  3. deep_dive_analysis: Multi-stage investigations involving hypothesis formulation, data join assemblies, multi-query execution, and visual charting.
  • Analyses: Explore the primary analytical container and interface where Kowalski is deployed.
  • Deep Dives: Learn how Kowalski plans and executes autonomous, multi-step agent workflows.
  • Knowledge Base: See how custom context documents are injected to ground Kowalski's analytical runs.
  • OLAP (DuckDB): Understand the database architecture Kowalski queries.