Code Routes Documentation

This document describes the API endpoints available for code execution, editing, fixing, and cleaning operations in the Auto-Analyst backend.

Base URL

All code-related endpoints are prefixed with /code.

Endpoints

Execute Code

Executes Python code against the current session's dataframe.

Endpoint: POST /code/execute

Request Body:

{
    "code": "string",              // Python code to execute
    "session_id": "string",        // Optional session ID
    "message_id": 123              // Optional message ID for tracking
}

Response:

{
    "output": "string",            // Execution output
    "plotly_outputs": [            // Optional array of plotly outputs
        "string"
    ]
}

Error Responses:

400 Bad Request: No dataset loaded or no code provided
500 Internal Server Error: Execution error

Edit Code

Uses AI to edit code based on user instructions.

Endpoint: POST /code/edit

Request Body:

{
    "original_code": "string",     // Code to be edited
    "user_prompt": "string"        // Instructions for editing
}

Response:

{
    "edited_code": "string"        // The edited code
}

Error Responses:

400 Bad Request: Missing original code or editing instructions
500 Internal Server Error: Editing error

Fix Code

Uses AI to fix code with errors, employing a block-by-block approach with DSPy refinement.

Endpoint: POST /code/fix

Request Body:

{
    "code": "string",              // Code containing errors
    "error": "string"              // Error message to fix
}

Response:

{
    "fixed_code": "string"         // The fixed code
}

Error Responses:

400 Bad Request: Missing code or error message
500 Internal Server Error: Fixing error

Clean Code

Cleans and formats code by organizing imports and ensuring proper code block formatting.

Endpoint: POST /code/clean-code

Request Body:

{
    "code": "string"               // Code to clean
}

Response:

{
    "cleaned_code": "string"       // The cleaned code
}

Error Responses:

400 Bad Request: No code provided
500 Internal Server Error: Cleaning error

Get Latest Code

Retrieves the latest code from a specific message.

Endpoint: POST /code/get-latest-code

Request Body:

{
    "message_id": 123              // Message ID to retrieve code from
}

Response:

{
    "code": "string"               // The retrieved code
}

Error Responses:

400 Bad Request: Missing message ID
404 Not Found: Message not found
500 Internal Server Error: Retrieval error

Code Processing Features

Import Organization

The code processing system automatically:

Moves all import statements to the top of the file
Deduplicates imports
Sorts imports alphabetically

Code Block Management

The system supports code blocks marked with special comments:

Start marker: # agent_name code start
End marker: # agent_name code end

Error Handling with DSPy Refinement

When fixing code, the system uses DSPy's refinement mechanism:

Identifies specific code blocks with errors
Processes error messages to extract relevant information
Uses a scoring function to validate fixes
Employs iterative refinement with up to 3 attempts
Fixes each block individually while maintaining the overall structure
Preserves code block markers and relationships

Dataset Context

When editing or fixing code, the system provides context about the current dataset including:

Number of rows and columns
Column names and data types
Null value counts
Sample values for each column

Code Execution Safety

The execution system includes safety measures:

Removes blocking calls like plt.show()
Handles __main__ block extraction
Cleans up print statements with unwanted newlines
Executes code in isolated namespaces

Session Management

All endpoints require a valid session ID, which is used to:

Access the current dataset
Maintain state between requests
Track code execution history
Store execution results for analysis

Error Handling

The system provides detailed error messages while maintaining security by:

Logging errors for debugging
Returning user-friendly error messages
Preserving original code in case of processing failures
Using code scoring to validate fixes before returning results