Vectora.ai Workflow Documentation

This document outlines the core workflows and feature implementations for the Vectora.ai application, focusing on Data Analysis, Visualization, and Advanced Data Cleaning.

1. Authentication & Session Management

• Flow: Users sign in via /signin.
• Persistence:
  • Login state matches the username stored in localStorage.
  • The application automatically checks this state on load.
  • If the user is not logged in, they are redirected or prompted to sign in.
  • Fix Implemented: Removed auto-logout timers to ensure sessions remain active as long as the user wants.

2. Data Upload & Persistence

• Flow: Users upload a CSV file on the Landing Page (/).
• Caching:
  • Before uploading to the server, the file (if < 5MB) is saved to the browser's localStorage (cached_csv).
  • This ensures that if the page is refreshed or the browser closed, the analysis session can be restored immediately without re-uploading.
• Server State: The server stores the dataframe in an in-memory global state (server/state.py).

3. Intelligent Data Analysis (Chat)

• Flow: Users interact with their data via natural language on the Chat Page (/chat).
• Process:
  1. Intent Detection: The system detects if the user wants to visualize, clean, OR analyze data based on keywords (e.g., "plot", "chart", "clean").
  2. Plan Generation: An LLM (GPT-4o) converts the textual question into a structured JSON execution plan.
  3. Execution: The plan is executed against the Pandas DataFrame.
  4. Response: The system generates a natural language explanation of the results.
• Visualization:
  • If the user asks to "visualize" or toggles the visualization checkbox, the system prioritizes sales_diagnostics or chart-compatible operators.
  • Interactive charts are generated and displayed in the chat stream.

4. Advanced Data Cleaning Studio

• Flow: Users access the dedicated cleaning interface at /clean.
• Modes: The system supports problem-specific cleaning pipelines:
  • General: Standard handling of missing values (mean/mode imputation) and duplicate removal.
  • Sentiment Analysis (NLP): Optimized for text data. Automatically detects text columns, performs lowercasing, punctuation removal, and whitespace normalization.
  • Binary Classification: Detects target class imbalance. If found (e.g., 90% vs 10%), it automatically performs undersampling to balance the dataset for better model training.
  • Classification: Analyzes and reports class distributions for multi-class problems.
• Reporting:
  • Instead of a black box, the system returns a detailed "Analyst Report" summarizing exactly what actions were taken (e.g., "Removed 15 duplicates", "Balanced dataset from 1000 to 400 rows").