0.1.1

Version 0.1.1

New Optimization - Quick Fixes

I have introduced a "Quick Fix" mechanism to significantly reduce token usage and context size:

• Structured Analysis:
• • error_analysis.py now provides a JSON response that categorizes the error.
• • Missing Imports Logic: If the AI detects a missing library (e.g., numpy), executor.py now prepends the import statement directly to the existing code and retries execution.
• • Bypassing Regeneration: This bypasses the expensive CodeAdaptationAgent fix_code call, saving hundreds of tokens and preventing the AI from accidentally changing working parts of the code while trying to fix a simple import. This makes the system much more efficient and robust against context window limits.

QLX AI powered data science by Qualex

Engineer: Antonio Martínez

0.1.0

Version 0.1.0

Check latest Version

Have implemented limits on the error message size and the history of failed attempts to ensure that the LLM context is not exceeded when the AI tries to fix code multiple times.

Changes Made

• 1. Truncating stderr in ExecutorService
     I modified executor.py to truncate the standard error output (stderr) to the last 2000 characters before sending it to the AI for analysis or fixing. Usually, the most relevant traceback info is at the end of the error log.
• 2. Limiting Error History Storage
     In executor.py, the error_history list now only stores the last 500 characters of each error message to keep the object size manageable.
• 3. Pruning History in CodeAdaptationAgent
     In code_adaptor.py, the fix_code method now only includes the last 3 failed attempts in its prompt. This prevents the "history section" of the prompt from growing indefinitely.

QLX AI powered data science by Qualex

Engineer: Antonio Martínez

0.0.9

Version 0.0.9 (deprecated)

Check latest Version

I have fixed the issue where long-running ML training tasks (when retries>=3, like the one with 49 Random Forest trials) would cause a "network error". This happened because the server was not sending any data while waiting for the training to finish, leading to a timeout.

Changes Made

• 1. Asynchronous Subprocess Streaming
     Updated the ExecutorService (executor.py) to run pipeline scripts using asynchronous subprocesses.
     Instead of waiting for the script to finish, it now reads the output line by line. Each line is sent immediately to the frontend as a status update.

This keeps the network connection active and provides real-time feedback.

• 2. Verbose ML Templates
     Updated the ML templates to provide more progress information during training:
• • • AutoML: Now prints the current algorithm being optimized.
• • • Random Forest: Now prints when SHAP value calculation starts.
• • • Scikit-Learn: Enabled verbose=1 in RandomizedSearchCV to show trial progress in the logs.

QLX AI powered data science by Qualex

Engineer: Antonio Martínez

0.0.8

Version 0.0.8 (deprecated)

Check latest Version

Recreated graphic creations with plotly

QLX AI powered data science by Qualex

Bug found: Executor has an error when retries >3

Engineer: Antonio Martínez

0.0.7

Version 0.0.7 (deprecated)

Check latest Version

Recreated graphic creations with plotly

QLX AI powered data science by Qualex

Engineer: Antonio Martínez

0.0.6

Version 0.0.6 (deprecated)

Added very nice stuff and a sandbox to test previous trained ML models

Check latest Version

0.0.5

Version 0.0.5 (deprecated)

Check latest Version

Changes Made

1. User Interface Enhancement
   In the Advanced Analysis page, I added a new section "Project Objective" with a text area. This is where you can describe your goal (e.g., "Predict customer churn").

Advanced Analysis UI (Note: Placeholder caption for UI)

2. Objective-Driven SQL & EDA
   Dynamic Dataset Construction: When an objective is provided, the Automatic EDA Agent now uses the LLM to write a custom SQL query. This query can join multiple tables to build exactly the dataset needed for your goal.
   Targeted EDA: All subsequent EDA phases (Statistics, Missing Data, Correlations) now take the objective into account when generating insights.

3. Tailored Machine Learning Code
   The Code Adaptation Agent uses the ML objective to customize the load_data and preprocessing steps in the final Python script.
   The Insights Agent now evaluates the model's performance specifically against the stated objective.

4. Improved Join Accuracy & Schema Context
   AI-Friendly Schema: Added a specialized text-based schema representation (get_llm_schema_context) that explicitly lists columns and foreign key relationships.

Strict SQL Rules & Prompt Sanitization:
Removed all specific examples (like casino_id) from AI prompts to prevent "mimicry" of non-existent columns.
Added a Mandatory Reasoning Step: Agents must now list and verify columns from the schema BEFORE writing any code.
Improved Error Fixing: Added logic to prevent agents from repeating failed join keys, forcing them to find alternatives or simplify the query.

5. Technical Implementation
   Backend Endpoints: Updated Pydantic models in endpoints.py to support ml_objective.
   API Client: Updated client.ts to pass the new parameter through all stages of the pipeline.
   State Management: Ensured the objective is preserved across navigation between AdvancedAnalysis, EDAProgress, and Results pages.

0.0.4

Version 0.0.4

Enabled selection of tables in advanced mode

Check latest Version

0.0.3

Agentic SQL mode in the manual EAD mode

Check latest Version

0.0.2

Version 0.0.2

Check latest Version