Business Analytics Agent
Ask questions in plain language and receive SQL-backed charts and business insights.
The evidence shows restrained GitHub workflow permissions by default, with contents: write confined to the release job. It also states that Feishu writes occur only after an explicit member request, that remote-GPU passwords use OS credential storage, and that authorship and licensing are identified. Deductions apply because these safeguards cover only selected paths: there is no comprehensive account of how uploaded data, database contents, knowledge-base material, long-term memory, third-party LLMs, or MCP servers receive and retain data; general confirmation is not demonstrated for SQL execution, export, remote MCP, or other effects; dependencies mostly use broad minimum ranges and the referenced production lockfile is not supplied; and no rollback mechanism is documented for database or external writes. No red-line behavior is evidenced.
The sources provide substantial installation, runtime-entry, platform-dependency, and common-failure guidance. The chart-selector tests also verify that an unmatched request produces a clarification path, supporting ordinary dependency availability and failure messaging. Deductions reflect internal inconsistencies: the README calls 1.3.0 LTS current while SECURITY.md supports 1.4.0 LTS, the Windows installer example remains at 1.2.0, and the badge says 41 chart types while the capability section says 43. The release test job compiles runtime sources but does not show the full test suite running across its release matrix.
The product clearly targets nontechnical business-analysis users and documents concrete scenarios involving sales trends, regional profit, charts, clustering, decision trees, exports, and several data sources. Slash commands, provider configuration, source lists, and a test for clarifying unmatched chart requests provide useful evidence for boundaries and triggering. Deductions apply because natural-language intent rules, ambiguous or hazardous SQL, read-versus-write boundaries, and MCP authorization are not systematically specified. Cross-platform coverage is good but remains complicated by database drivers, credential-store behavior, frontend Node requirements, and packaging-specific environments.
The README has a strong navigable structure covering capabilities, installation, commands, examples, FAQ, contribution, licensing, and changes. It also discloses some limitations, including planned GPU support, the extra SQL Server driver, and unsigned or unnotarized release artifacts. Deductions apply because naming varies among Chinese and English product names and the frontend package name, while versions 1.2, 1.3, and 1.4 appear inconsistently. The license clearly expresses a noncommercial dual-licensing intent and links the full CC terms, but combines custom wording with CC BY-NC 4.0, which is not a conventional software license. An email, Issues path, private vulnerability channel, and response commitments identify an update path, though responsibility remains centered on one named, externally unverified publisher.
The intended outputs—SQL, charts, business insights, Excel, DOCX, PPT, and streaming progress—are highly usable for the stated audience, and the integrated conversational workflow offers plausible value over manual SQL and disconnected analysis tools. Deductions apply because most effectiveness evidence consists of README claims and depicted interfaces; the supplied test covers chart selection only and cannot statically substantiate insight quality or the complete workflow. Cost discussion covers initial setup time, disk space, and the lighter ONNX choice, but does not quantify LLM charges, compute consumption, privacy costs, or performance on large datasets.
Some README claims are corroborated by package.json, dependency declarations, the build workflow, and chart-selection tests, including the stack, entry points, build process, and clarification behavior. Planned GPU support, unsigned release packages, and the pending copyright registration number are appropriately qualified, providing some separation between current facts and future status. Deductions apply because claims about 43 charts, automatic SQL and insights, complete source support, improved long-term memory, and response timelines lack supplied implementation excerpts, targeted tests, or reports for direct tracing. Version and chart-count conflicts further weaken cross-source verification.
- Before connecting real business data, determine whether uploads, query results, knowledge-base content, and long-term memory are sent to the configured LLM or remote MCP services, and establish retention and deletion rules for every recipient.
- Do not treat natural-language-generated SQL as inherently safe. Use a read-only database account, inspect generated statements, and require explicit confirmation and backups before any write or schema change.
- The one-line installers download and execute network content directly. Prefer a revision-pinned download, review the script, and verify release artifacts. The workflow explicitly describes desktop packages as unsigned or unnotarized.
- Most dependency declarations are minimum-version ranges, and requirements.lock.txt is not included in the evidence. Production use should rely on audited locks and ongoing vulnerability scanning.
- Version status is inconsistent across the README, installer example, and security policy, which reference 1.2, 1.3, and 1.4. Confirm which supported release actually corresponds to the assessed revision.
- CC BY-NC 4.0 and the additional commercial restrictions may be unsuitable for enterprise or commercial deployment. Obtain explicit authorization and legal review before such use.
What does this agent do, and when should you use it?
Business Analytics Agent is a self-hosted web application for conversational business analysis, with a Flask backend and SSE streaming for live progress. Users can upload Excel or CSV files or connect SQLite, MySQL, PostgreSQL, and SQL Server databases, then query the data in natural language. The application inspects the schema, generates and runs SQL, recommends a visualization from 43 chart types across six categories, and returns an analytical summary. Its broader toolkit includes outlier treatment, decile analysis, K-Means clustering, decision trees, and exports to Excel, DOCX, and presentation formats. It runs locally on Windows, macOS, or Linux, while model APIs and remote data services may remain outside the local deployment boundary.
After a data source is connected, the application interprets the question, reads the Schema, generates SQL, executes the query, selects a chart, and writes business insights; SSE streams the stages back to the interface. It reads Excel and CSV files and queries SQLite, MySQL, PostgreSQL, and SQL Server. The v1.3.0 notes also describe a Feishu bot flow that reads Bitable records, analyzes them through DuckDB/SQL, and posts conclusions to group chat, creating or modifying result records only after an explicit member request. Its command surface includes /chart, /sql, /analyze, /tree, /kmeans, /data, /inset, /winsorize, /trimming, /export, /report, /ppt, and /status. Chart HTML files are retained under outputs/charts, while processed spreadsheets, DOCX reports, and built-in-style presentations can be exported. Model configuration supports DeepSeek, OpenAI, AtlasCloud, and arbitrary OpenAI SDK-compatible APIs through configurable base_url, model, and api_key values. The application can also connect local or remote MCP services and accept uploaded business knowledge.
- A sales or operations user who does not write SQL uploads a spreadsheet and asks for the last 12 months of sales trends, receiving a query, line chart, and growth commentary.
- A regional manager connects an operational database and asks which region has the highest profit, producing a ranking, bar chart, and regional insights.
- A data analyst uses
/winsorize,/trimming, decile grouping,/kmeans, or/treefor cleaning, segmentation, and exploratory modeling. - A reporting team uses
/export,/report, and/pptto turn completed analysis into processed data, a document, or a presentation. - A team working from Feishu Bitable asks the group bot to analyze records and, when explicitly requested, create a result table or append and update records.
- An organization that prefers local application hosting runs the Flask service on its own machine while choosing its own OpenAI-compatible model endpoint, databases, knowledge base, and MCP connections.
What are this agent's strengths and limitations?
- It covers the complete path from schema inspection and SQL generation through query execution, chart recommendation, and written insights, with SSE exposing progress.
- Its documented inputs include Excel, CSV, and four database systems; v1.3.0 also adds a Feishu bot and Bitable analysis loop.
- The visualization catalog is unusually broad for a conversational analyst, listing 43 chart types across comparison, time, distribution, geographic, relationship, and part-to-whole categories.
- The tool extends beyond question answering with outlier treatment, decile analysis, K-Means, decision trees, and Excel, DOCX, and presentation exports.
- Model access is not limited to one named provider: DeepSeek, OpenAI, AtlasCloud, and custom OpenAI SDK-compatible endpoints are documented.
- Deployment requires Python 3.10 or later, and the first dependency installation is documented to take roughly 3–5 minutes; hosted model APIs and remote services also require connectivity and credentials.
- SQL Server setup has an additional system dependency on Microsoft ODBC Driver 17 or 18 and may require TCP/IP, port, firewall, encryption, and permission troubleshooting.
- The README declares CC BY-NC 4.0 and prohibits unauthorized commercial use, while the supplied repository metadata reports NOASSERTION; commercial adopters should verify the LICENSE file and obtain authorization.
- The general data-source section still labels DuckDB and Spark as future work. The Feishu-specific DuckDB flow in v1.3.0 is not evidence of complete general-purpose connectors for both systems.
- GPU compute is explicitly marked as planned rather than generally available pending end-to-end validation.
How do you install or deploy this agent?
Python 3.10 or later is required. A source installation is documented as:
git clone https://github.com/Zafer-Liu/Data-Analysis-Agent.git
cd Data-Analysis-Agent
pip install -r requirements.txt
python app.pyOpen http://localhost:5001, then enter the API Key, Base URL, and Model in the sidebar settings. Windows 10/11 64-bit users may alternatively use the release installer. From the archive distribution, Windows users can run start.bat; on macOS, run chmod +x ~/Downloads/Data-Analysis-Agent/start.command and xattr -d com.apple.quarantine ~/Downloads/Data-Analysis-Agent/start.command, then open start.command. The documented online installers are iwr -useb https://raw.githubusercontent.com/Zafer-Liu/Data-Analysis-Agent/main/install.ps1 | iex in Windows PowerShell and curl -fsSL https://raw.githubusercontent.com/Zafer-Liu/Data-Analysis-Agent/main/install.sh | sh on macOS/Linux. The first launch creates a virtual environment and installs dependencies, with a documented estimate of 3–5 minutes.
How do you use this agent?
Run python app.py and visit http://localhost:5001. Configure an API Key, Base URL, and Model first; the documented defaults are DeepSeek deepseek-v4-flash, OpenAI gpt-4o-mini, and AtlasCloud deepseek-v4-pro. Upload an Excel/CSV file or configure a database connection, then ask a question such as “Show sales trends for the last 12 months.” Use /chart user growth to prioritize a visualization, /sql to execute SQL directly, or /analyze, /tree, and /kmeans for deeper analysis. MySQL uses mysql+pymysql://username:password@host:port/database. SQL Server additionally requires Microsoft ODBC Driver 17 or 18; a Driver 18 example is mssql+pyodbc://username:password@server:1433/database?driver=ODBC+Driver+18+for+SQL+Server&TrustServerCertificate=yes. Open generated chart HTML from outputs/charts, or invoke /export, /report, or /ppt for deliverable files.
How does this agent compare with similar options?
The repository contrasts its workflow with traditional BI tools: users converse with the data while the application handles schema inspection, SQL generation, execution, visualization selection, and streaming progress. This can reduce manual query and report construction, but the supplied material contains no benchmark, pricing study, or feature matrix against a named BI product.
FAQ
Can it operate entirely offline?
Is commercial use permitted without payment?
What should I check if SQL Server will not connect?
pyodbc, enable SQL Server TCP/IP, verify the actual host and port, allow the port through the firewall, and confirm the account has database access.Are generated charts retained after restart?
outputs/charts; the corresponding HTML files can be opened directly in a browser.Which model services are supported?
base_url, model, and api_key.