🌌 DataVerse AI

An AI Data Scientist backend — parse, analyze, predict & explain your data.

---

DataVerse AI is a clean, production-ready MVP for an AI Data Scientist backend using FastAPI. It parses datasets (CSV/XLSX), validates columns, normalizes headers, computes business metrics deterministically using Pandas/scikit-learn, and produces a professional data report.

The LLM is optional and only used to polish the narration of computed facts. All calculations are deterministic.

Two-Agent Architecture

The MVP relies on exactly two agents with a clear division of labor:

graph TD
    User([User]) -->|1. Uploads CSV/XLSX| DatasetAgent[DatasetAgent]
    DatasetAgent -->|2. Validates & Parses| DatasetAgent
    DatasetAgent -->|3. Profiles Columns & Quality Scan| DatasetAgent
    DatasetAgent -->|4. Persists Session| Store[(Session Store)]
    DatasetAgent -->|5. Delivers Profile & Quality| Workspace[Analyze Workspace]
    
    Workspace -->|6. Triggers Analysis| AnalystAgent[AnalystAgent]
    AnalystAgent -->|7. Semantic KPI Mapping| AnalystAgent
    AnalystAgent -->|8. Runs EDA & AutoML Modeling| AnalystAgent
    AnalystAgent -->|9. SHAP Explainable AI XAI| AnalystAgent
    AnalystAgent -->|10. Compiles Report max 2 charts| ReportCompiler[Report Compiler]
    
    ReportCompiler -->|11. Generates Compact HTML & PDF| Store
    ReportCompiler -->|12. Renders Preview & XAI Card| ReportPage[Report & XAI Page]

Loading

Division of Labor

1. DatasetAgent (app/agents/dataset_agent.py):
   • Validates uploaded file limits and file formats.
   • Parses CSV/XLSX safely.
   • Normalizes columns (headers are cleaned and whitespace is removed).
   • Generates a unique session_id.
   • Stores the dataset locally in the filesystem session store.
   • Produces a dataset profile and data quality summary.
2. AnalystAgent (app/agents/analyst_agent.py):
   • Understands user semantic queries.
   • Maps columns semantically (e.g. mapping date, products, revenue).
   • Computes business metrics, EDA, trends, correlations, and outlier flags using Pandas.
   • Triggers predictive machine learning (Ridge or RandomForest) only if the dataset has at least 30 rows (MIN_ROWS_FOR_PREDICTION) and a target is provided.
   • Runs XAI (Shapley/Feature Importance) upon successful modeling.
   • Generates charts-ready JSON and final polished report narration (using offline deterministic narration if LLM keys are absent).

---

Unique Capabilities

Beyond the verifiable-analyst core (provenance receipts, reproducibility certificate, verified what-if, quality doctor), DataVerse AI ships three capabilities no comparable tool combines:

1. Root-Cause Investigator — ask "Why did revenue drop in May?" and the system runs a deterministic multi-step investigation: it locates the period delta, decomposes the change across every dimension (product / category / region / customer), ranks drivers by contribution ("Widget explains 100% of the drop"), and splits the change into price vs volume effects. Every step carries a provenance receipt. Works fully offline. (app/services/root_cause.py, POST /api/sessions/{sid}/datasets/{did}/investigate)
2. Counterfactual XAI — for each explained prediction, a deterministic single-feature search finds the smallest change that flips the outcome: "Raising quantity from 8 to 9.2 (+15%) would flip the predicted label from 'low' to 'high'." Goes beyond SHAP importance to actionable, reproducible explanations. (app/services/counterfactual.py)
3. Agentic chat loop — when an LLM key is configured, chat answers come from a real plan→act→observe agent: the LLM chooses deterministic tools (KPIs, segments, trends, what-if, root-cause, prediction/XAI, quality), reads the computed observations, and answers using only those numbers. The full tool trace streams live into the UI. With no key, it falls back to the deterministic answer path. (app/services/agent_loop.py)

All three preserve the core guarantee: every number is computed in pandas/scikit-learn; the LLM only plans and phrases.

---

Simple Setup

1. Configure Environment

Create a local .env file inside the dataverse_backend folder:

copy dataverse_backend\.env.example dataverse_backend\.env

2. Install MVP Requirements

Ensure your virtual environment is active, then install the lightweight MVP dependencies:

.\.venv\Scripts\python -m pip install -r dataverse_backend/requirements-mvp.txt

3. Run Backend Server

Start the FastAPI server from the dataverse_backend directory:

cd dataverse_backend
python -m uvicorn app.main:app --reload --host 127.0.0.1 --port 8000

Or run directly from the workspace root directory:

python -m uvicorn app.main:app --reload --app-dir dataverse_backend --host 127.0.0.1 --port 8000

---

API Endpoints

The frontend uses the session-based API flow:

• GET /health/live - backend liveness check.
• GET /api/health - API health check.
• POST /api/sessions - create a chat session.
• POST /api/sessions/{session_id}/datasets/upload?auto_analyze=true - upload a dataset into the session.
• POST /api/sessions/{session_id}/analyze - run full analysis for a session dataset.
• POST /api/sessions/{session_id}/messages - ask follow-up questions using content and dataset_id.
• GET /api/sessions/{session_id} - load messages, datasets, agent runs, and reports.
• GET /api/datasets - list recent datasets for the current workspace.

---

Testing with Curl

1. Create a Session

curl.exe -X POST http://localhost:8000/api/sessions `
  -H "Content-Type: application/json" `
  -d "{\"title\":\"New Chat\"}"

2. Upload and Auto-Analyze a Dataset

curl.exe -X POST "http://localhost:8000/api/sessions/YOUR_SESSION_ID/datasets/upload?auto_analyze=true" `
  -F "file=@sample_sales.csv"

3. Ask a Follow-Up Question

curl.exe -X POST http://localhost:8000/api/sessions/YOUR_SESSION_ID/messages `
  -H "Content-Type: application/json" `
  -d "{\"content\":\"examine it\",\"dataset_id\":\"YOUR_DATASET_ID\"}"

---

Expected Response Shapes

POST /api/analyze/upload Response

{
  "session_id": "0bde26cd-21cd-413b-bf64-b968ee631007",
  "filename": "sample_sales.csv",
  "dataset_profile": {
    "row_count": 40,
    "column_count": 6,
    "columns": ["date", "product", "category", "quantity", "revenue", "cost"],
    "dtypes": {"date": "object", "product": "object", "category": "object", "quantity": "int64", "revenue": "int64", "cost": "int64"}
  },
  "data_quality": {
    "data_quality_score": 1.0,
    "missing_cells": 0,
    "duplicate_rows": 0,
    "warnings": []
  },
  "semantic_map": {
    "dataset_type": "transaction_ledger",
    "column_roles": {
      "date": "timestamp",
      "product": "category",
      "category": "category",
      "quantity": "quantity",
      "revenue": "revenue",
      "cost": "cost"
    }
  },
  "business_metrics": {
    "total_revenue": 21630,
    "total_profit": 11330,
    "gross_margin": 0.5238
  },
  "query_answer": {
    "answer": "Dataset uploaded and analyzed.",
    "facts": {}
  },
  "eda": {
    "summary": {
      "quantity": {"mean": 6.8, "min": 2, "max": 15},
      "revenue": {"mean": 472.5, "min": 100, "max": 1200}
    }
  },
  "trends": {
    "series": [
      {"value_column": "revenue", "direction": "upward", "slope": 3.4}
    ]
  },
  "correlations": {
    "strong_pairs": [
      {"column_a": "quantity", "column_b": "revenue", "correlation": 0.98}
    ]
  },
  "outliers": {
    "total_outlier_cells": 0
  },
  "prediction": {
    "status": "complete",
    "task_type": "regression",
    "target_column": "revenue",
    "selected_model": "Ridge",
    "test_metrics": {"rmse": 12.34, "r2": 0.99},
    "predictions_sample": []
  },
  "xai": {
    "status": "complete",
    "plain_english_explanation": "Quantity is the strongest driver of Revenue..."
  },
  "charts": [
    {"type": "line", "title": "Sales revenue by month", "x": "period", "y": "sales_revenue", "data": []}
  ],
  "executive_summary": "Dataset contains 40 rows. Total revenue is 21630...",
  "key_insights": [
    "Dataset contains 40 rows and 6 columns.",
    "Total revenue is 21630."
  ],
  "recommendations": [
    "Review missing values before operational decisions."
  ],
  "warnings": [],
  "next_questions": [
    "Which target column should be optimized next?"
  ]
}

---

Verification Tests

Run the full end-to-end test suite:

cd dataverse_backend
..\.venv\Scripts\python -m pytest -v tests/test_mvp_e2e.py

All 10 test scenarios are validated and pass successfully in the local execution context.