VerbaGPT is an LLM tool for data analysis scenarios, with the core goal of truly applying cutting-edge large language models to data work. It allows users to connect to databases, ask questions in natural language, and directly obtain charts, insights, and analysis results without writing SQL or Python.
Purpose and Positioning
VerbaGPT addresses three bottlenecks organizations face in data utilization: the high SQL barrier, LLMs' lack of data context, and users' poor prompting skills. It tackles these issues through natural language to Python/SQL, a data annotation layer, and prompt libraries. It is officially called Text-to-Python, not just Text-to-SQL, so it supports tasks that are difficult to accomplish with SQL, such as statistical modeling, machine learning, and custom visualizations.
Main Capabilities
- Natural language queries: Ask questions in English, such as "What were the top 10 products last quarter," and get instant results and charts.
- Multi-model support: Virgo can orchestrate cutting-edge models such as GPT-4 and Claude, used respectively for schema reasoning, code generation, and review.
- Files and databases: Supports uploading CSV, Excel, and PDF files, or importing from Google Drive, while also connecting to live databases.
- Advanced analytics: Regression, time series forecasting, hypothesis testing, classification, clustering, decision trees, and custom charts such as heatmaps and 3D plots.
- Data Notes: Describe database structure, table relationships, column meanings, and business logic in natural language; the LLM automatically obtains context and generates accurate queries.
- Prompt Libraries: Package effective prompts into reusable collections, supporting sharing within an organization or licensing to customers.
- Footnotes and inspiration: Prompts shared by colleagues are presented through semantic search, helping new members get up to speed quickly.
- Local connections: Query local or LAN databases through the Taurus connector without exposing them to the internet; connection strings and passwords remain local, and access is read-only.
- Private inference: Run models locally for writing and reviewing analysis code through Ollama, LM Studio, or any OpenAI-compatible engine; if the local connection is lost, it will clearly indicate this and will not silently fall back to a cloud model.
- Taurus agent: A local agent based on Claude Code, with full system permissions and file access, suitable for sensitive data, complex statistical modeling, and exploratory data science.
- Customer access: Consulting firms can publish databases and invite customer emails, and customers do not needSubscribe or pay per seat, use the organization credit pool, set credit limits and revoke at any time.
- Learning capability: workspace private memory, can save corrections as rules, reuse successful methods, and generate daily data annotation improvement drafts for review.
How to use
Users can use cloud analysis through the web app app.verbagpt.com, or run the Taurus local agent. Basic workflow: ask a question in natural language → Virgo uses embeddings to find relevant schema columns → generates Python code with SQL queries → executes safely in a sandbox environment → returns charts and tables. Taurus workflow: receive the question → plan a multi-step approach (inspect schema, query data, analyze, visualize) → iteratively execute tools (Python REPL, SQL, schema inspector) → self-correct → return the complete analysis.
Use cases
- Standard analytical queries and team collaboration (shared chat history)
- Production-grade web deployment, multi-user support
- Complex statistical modeling and exploratory data science
- Process local files and databases, keeping data local
- Consulting firms providing database question-and-answer access to clients
- Sensitive data work requiring private inference or a local agent
Limitations and notes
- Using local network databases or private inference requires running the Taurus application; these features are included in all plans.
- In cloud mode, query results, questions, data annotations, and schema selections pass through VerbaGPT; in Taurus mode, only questions and printed results are sent to the model.
- Database credentials are not sent to the AI model.
- Supports PostgreSQL, MySQL, and SQL Server, including Windows login.
- Security aspects: JWT authentication, bcrypt password hashing, role-based permissions, AES-256 encryption at rest, TLS 1.3 encryption in transit, isolated code execution, audit logs (30-day retention), 99.9% uptime SLA.
- The official statement is that chat history or data is not used to train AI models.
- Specific details such as pricing, free quota, and platform availability are not yet provided.
