llm-accounting 0.1.5

A package for tracking LLM usage and costs

LLM Accounting

A Python package for tracking and analyzing LLM usage across different models and applications. It is primarily designed as a library for integration into development process of LLM-based agentic workflow tooling, providing robust tracking capabilities. While its main use is as a library, it also provides a powerful CLI for scripting and batch workloads.

Keywords: LLM, accounting, usage tracking, cost management, token counting, agentic workflows, AI, Python

Features

• Track usage of different LLM models
• Record token counts (prompt, completion, total)
• Track costs and execution times
• Support for local token counting
• Pluggable backend system (SQLite included, Neon/PostgreSQL fully supported)
• CLI interface for viewing and tracking usage statistics
• Support for tracking caller application and username
• Automatic database schema migration (for supported backends)
• Strict model name validation
• Automatic timestamp handling
• Comprehensive audit logging for all LLM interactions

Installation

pip install llm-accounting

For specific database backends, install the corresponding optional dependencies:

# For SQLite (default)
pip install llm-accounting[sqlite]

# For Neon/PostgreSQL
pip install llm-accounting[neon]

Usage

Basic Usage

from llm_accounting import LLMAccounting
# from llm_accounting.backends.sqlite import SQLiteBackend # Default
# from llm_accounting.backends.neon import NeonBackend # If using Neon
# from datetime import datetime # If providing timestamps or querying by date

# Default backend (SQLite)
# with LLMAccounting() as accounting:
#     # Track usage (model name is required, timestamp is optional)
#     accounting.track_usage(
#         model="gpt-4",  # Required: name of the LLM model
#         prompt_tokens=100,
#         completion_tokens=50,
#         total_tokens=150,
#         cost=0.002,
#         execution_time=1.5,
#         caller_name="my_app",  # Optional: name of the calling application
#         username="john_doe",   # Optional: name of the user
#         timestamp=None         # Optional: if None, current time will be used
#     )
#     
#     # Get statistics
#     # start_date = datetime(2024, 1, 1)
#     # end_date = datetime(2024, 1, 31)
#     # stats = accounting.get_period_stats(start_date, end_date)
#     # model_stats = accounting.get_model_stats(start_date, end_date)
#     # rankings = accounting.get_model_rankings(start_date, end_date)

Note: The LLMAccounting class and its methods are synchronous. If you are integrating llm-accounting into an asynchronous application, you should run its synchronous calls in a separate thread (e.g., using asyncio.to_thread) to avoid blocking the event loop.

CLI Usage

# Track a new usage entry (model name is required, timestamp is optional)
llm-accounting track \
    --model gpt-4 \
    --prompt-tokens 100 \
    --completion-tokens 50 \
    --total-tokens 150 \
    --cost 0.002 \
    --execution-time 1.5 \
    --caller-name my_app \
    --username john_doe \
    --timestamp "2024-01-01T12:00:00" \
    --cached-tokens 20 \
    --reasoning-tokens 10

# ... (other CLI examples remain the same) ...

# Show today's stats
llm-accounting stats --daily

# Show stats for a custom period
llm-accounting stats --start 2024-01-01 --end 2024-01-31

# Show most recent entries
llm-accounting tail

# Show last 5 entries
llm-accounting tail -n 5

# Delete all entries
llm-accounting purge

# Execute custom SQL queries (if backend supports it and it's enabled)
# llm-accounting select --query "SELECT model, COUNT(*) as count FROM accounting_entries GROUP BY model"

### Usage Limits

The `llm-accounting limits` command allows you to manage usage limits for your LLM interactions. You can set, list, and delete limits based on various scopes (global, model, user, caller) and types (requests, input tokens, output tokens, cost) over specified time intervals.

#### Set a Usage Limit

Set a new usage limit. For example, to set a global limit of 1000 requests per day:

```bash
llm-accounting limits set \
    --scope GLOBAL \
    --limit-type requests \
    --max-value 1000 \
    --interval-unit day \
    --interval-value 1

To set a cost limit of $5.00 per hour for a specific user:

llm-accounting limits set \
    --scope USER \
    --username john_doe \
    --limit-type cost \
    --max-value 5.00 \
    --interval-unit hour \
    --interval-value 1

To set an input token limit of 50000 tokens per week for a specific model:

llm-accounting limits set \
    --scope MODEL \
    --model gpt-4 \
    --limit-type input_tokens \
    --max-value 50000 \
    --interval-unit week \
    --interval-value 1

List Usage Limits

List all configured usage limits:

llm-accounting limits list

Delete a Usage Limit

Delete a usage limit by its ID (you can find the ID using llm-accounting limits list):

llm-accounting limits delete --id 1

Database Backend Selection via CLI

You can specify the database backend directly via the CLI using the --db-backend option. This allows you to switch between sqlite (default) and neon without modifying code.

# Use SQLite backend (default behavior, --db-backend can be omitted)
llm-accounting --db-backend sqlite --db-file my_sqlite_db.sqlite stats --daily

# Use Neon backend
# Requires NEON_CONNECTION_STRING environment variable to be set, or provide it directly
llm-accounting --db-backend neon --neon-connection-string "postgresql://user:[e-mail]/dbname?sslmode=require" stats --daily

# Example: Track usage with Neon backend
llm-accounting --db-backend neon \
    --neon-connection-string "postgresql://user:[e-mail]/dbname?sslmode=require" \
    track \
    --model gpt-4 \
    --prompt-tokens 10 \
    --cost 0.0001

### Shell Script Integration

The CLI can be easily integrated into shell scripts. Here's an example:

```bash
#!/bin/bash

# Track usage after an LLM API call
llm-accounting track \
    --model "gpt-4" \
    --prompt-tokens "$PROMPT_TOKENS" \
    --completion-tokens "$COMPLETION_TOKENS" \
    --total-tokens "$TOTAL_TOKENS" \
    --cost "$COST" \
    --execution-time "$EXECUTION_TIME" \
    --caller-name "my_script" \
    --username "$USER"

# Check daily usage
llm-accounting stats --daily

Database Schema

The database schema generally includes the following tables and key fields (specifics might vary slightly by backend, but NeonBackend adheres to this structure):

accounting_entries Table:

• id: SERIAL PRIMARY KEY - Unique identifier for the entry.
• model_name: VARCHAR(255) NOT NULL - Name of the LLM model.
• prompt_tokens: INTEGER - Number of tokens in the prompt.
• completion_tokens: INTEGER - Number of tokens in the completion.
• total_tokens: INTEGER - Total tokens (prompt + completion).
• local_prompt_tokens: INTEGER - Locally counted prompt tokens.
• local_completion_tokens: INTEGER - Locally counted completion tokens.
• local_total_tokens: INTEGER - Total locally counted tokens.
• cost: DOUBLE PRECISION NOT NULL - Cost of the API call.
• execution_time: DOUBLE PRECISION - Execution time in seconds.
• timestamp: TIMESTAMP WITHOUT TIME ZONE DEFAULT CURRENT_TIMESTAMP - Timestamp of the usage.
• caller_name: VARCHAR(255) - Optional identifier for the calling application/script.
• username: VARCHAR(255) - Optional identifier for the user.
• cached_tokens: INTEGER - Number of tokens retrieved from cache.
• reasoning_tokens: INTEGER - Number of tokens used for model reasoning/tool use.

usage_limits Table (for defining quotas/limits):

• id: SERIAL PRIMARY KEY
• scope: VARCHAR(50) NOT NULL (e.g., 'USER', 'GLOBAL')
• limit_type: VARCHAR(50) NOT NULL (e.g., 'COST', 'REQUESTS')
• max_value: DOUBLE PRECISION NOT NULL
• interval_unit: VARCHAR(50) NOT NULL (e.g., 'HOURLY', 'DAILY')
• interval_value: INTEGER NOT NULL
• model_name: VARCHAR(255) (Optional, for model-specific limits)
• username: VARCHAR(255) (Optional, for user-specific limits)
• caller_name: VARCHAR(255) (Optional, for caller-specific limits)
• created_at: TIMESTAMP WITHOUT TIME ZONE DEFAULT CURRENT_TIMESTAMP
• updated_at: TIMESTAMP WITHOUT TIME ZONE DEFAULT CURRENT_TIMESTAMP

Note: The id fields are managed internally by the database.

Backend Configuration

SQLite (Default)

The default backend is SQLite, which stores data in a local file. Below is a comprehensive example demonstrating how to configure a custom SQLite database file, track usage, set and check usage limits, and utilize the audit logger.

import os
from llm_accounting import LLMAccounting
from llm_accounting.backends.sqlite import SQLiteBackend
from llm_accounting.models.limits import LimitScope, LimitType, TimeInterval
import time
from datetime import datetime, timedelta
from llm_accounting.audit_log import AuditLogger

# Define custom database filenames
custom_accounting_db_filename = "my_custom_accounting.sqlite"
custom_audit_db_filename = "my_custom_audit.sqlite"

print(f"Initializing LLMAccounting with custom DB: {custom_accounting_db_filename}")

# 1. Initialize SQLiteBackend with the custom filename
sqlite_backend = SQLiteBackend(db_path=custom_accounting_db_filename)

# 2. Pass the custom backend to LLMAccounting
# Using a context manager ensures the connection is properly opened and closed
with LLMAccounting(backend=sqlite_backend) as accounting:
    print(f"LLMAccounting initialized. Actual DB path: {accounting.get_db_path()}")

    # Example usage: track some usage
    accounting.track_usage(
        model="gpt-4",
        prompt_tokens=100,
        completion_tokens=50,
        cost=0.01,
        username="example_user",
        caller_name="example_app"
    )
    print("Usage tracked successfully.")

    # Verify stats (optional)
    end_time = datetime.now()
    start_time = end_time - timedelta(days=1)
    stats = accounting.get_period_stats(start_time, end_time)
    print(f"Stats for last 24 hours: {stats.sum_cost:.4f} cost, {stats.sum_total_tokens} tokens")

    print("\n--- Testing Usage Limits ---")
    # Set a global limit: 10 requests per minute
    print("Setting a global limit: 10 requests per minute...")
    accounting.set_usage_limit(
        scope=LimitScope.GLOBAL,
        limit_type=LimitType.REQUESTS,
        max_value=10,
        interval_unit=TimeInterval.MINUTE,
        interval_value=1
    )
    print("Global limit set.")

    # Simulate requests and check quota
    for i in range(1, 15): # Try 14 requests to exceed the limit
        model = "gpt-3.5-turbo"
        username = "test_user"
        caller_name = "test_app"
        input_tokens = 10

        allowed, reason = accounting.check_quota(
            model=model,
            username=username,
            caller_name=caller_name,
            input_tokens=input_tokens
        )
        if allowed:
            print(f"Request {i}: ALLOWED. Tracking usage...")
            accounting.track_usage(
                model=model,
                prompt_tokens=input_tokens,
                cost=0.0001,
                username=username,
                caller_name=caller_name
            )
        else:
            print(f"Request {i}: DENIED. Reason: {reason}")
        
        # Small delay to simulate real-world requests, but not enough to reset minute limit
        time.sleep(0.1) 

print(f"\nInitializing AuditLogger with custom DB: {custom_audit_db_filename}")

# Initialize AuditLogger with the custom filename
with AuditLogger(db_path=custom_audit_db_filename) as audit_logger:
    print(f"AuditLogger initialized. Actual DB path: {audit_logger.get_db_path()}")

    # Example usage: log a prompt
    audit_logger.log_prompt(
        app_name="my_app",
        user_name="test_user",
        model="gpt-3.5-turbo",
        prompt_text="Hello, how are you?"
    )
    print("Prompt logged successfully.")

# Clean up the created database files (for example purposes)
print("\nCleaning up created database files...")
if os.path.exists(custom_accounting_db_filename):
    os.remove(custom_accounting_db_filename)
    print(f"Removed {custom_accounting_db_filename}")
if os.path.exists(custom_audit_db_filename):
    os.remove(custom_audit_db_filename)
    print(f"Removed {custom_audit_db_filename}")

print("\nExample complete.")

Neon Backend (PostgreSQL)

The NeonBackend provides a reference implementation for using a cloud-based PostgreSQL database with llm-accounting. It is specifically designed to work well with Neon serverless Postgres, but it can also be used with any other standard PostgreSQL instance.

1. Set Up Your Neon Database (User's Responsibility):

To use NeonBackend with Neon, you'll need to set up your own database instance:

• Sign Up: Go to https://neon.tech/ and sign up for an account. The free tier is suitable for experimentation and development.
• Create a Project: In the Neon console, create a new project. This will be your serverless Postgres instance.
• Obtain Connection String: Once the project is created, find your database's connection string (URI format). It will look something like this:

  postgresql://<user>:<password>@<host>.neon.tech:<port>/<dbname>?sslmode=require
  

  Note: Neon typically requires sslmode=require.

2. Install Dependencies:

The NeonBackend requires the psycopg2-binary package to communicate with PostgreSQL databases. You can install it as an extra dependency:

pip install llm-accounting[neon]

3. Configuration:

The NeonBackend primarily expects the database connection string to be available via the NEON_CONNECTION_STRING environment variable.

export NEON_CONNECTION_STRING="postgresql://your_user:your_password@your_host.neon.tech:5432/your_dbname?sslmode=require"

Replace the placeholder values with your actual Neon connection string.

Alternatively, if you are instantiating NeonBackend manually in your code, you can pass the connection string directly to its constructor (though using the environment variable is often preferred for flexibility).

4. Usage Example:

To use the NeonBackend, you need to instantiate it and pass it to the LLMAccounting class:

from llm_accounting import LLMAccounting
from llm_accounting.backends.neon import NeonBackend # Import the NeonBackend
# from datetime import datetime # if you are passing timestamps or querying by date

# Option 1: Connection string from environment variable NEON_CONNECTION_STRING
# Ensure NEON_CONNECTION_STRING is set in your environment before running the script.
# For example: export NEON_CONNECTION_STRING="your_neon_uri_here"

neon_backend_env = NeonBackend() # Reads from environment variable
accounting_neon_env = LLMAccounting(backend=neon_backend_env)

# The LLMAccounting class's methods might be async or sync.
# The NeonBackend's initialize/close methods are synchronous.
# This example assumes LLMAccounting manages the async/sync interaction if its methods are async.
# For simplicity, direct calls are shown here. If LLMAccounting methods are async, use `async with`.

# with accounting_neon_env:
#     # Example: Track usage
#     accounting_neon_env.track_usage(
#         model="gpt-3.5-turbo",
#         prompt_tokens=50,
#         completion_tokens=100,
#         cost=0.00015
#     )
#     print("Usage tracked with Neon backend (from env var).")
#     
#     # Example: Get stats for a period
#     # start_date = datetime(2024, 1, 1)
#     # end_date = datetime(2024, 1, 31)
#     # stats = accounting_neon_env.get_period_stats(start_date, end_date)
#     # print(stats)


# Option 2: Pass connection string directly
# Replace with your actual connection string if testing this way.
# neon_connection_str = "postgresql://user:[e-mail]/dbname?sslmode=require" 
# neon_backend_direct = NeonBackend(neon_connection_string=neon_connection_str)
# accounting_neon_direct = LLMAccounting(backend=neon_backend_direct)

# with accounting_neon_direct:
#     accounting_neon_direct.track_usage(
#         model="gpt-4",
#         prompt_tokens=200,
#         completion_tokens=400,
#         cost=0.006
#     )
#     print("Usage tracked with Neon backend (direct connection string).")

Error Handling/Notes:

• The NeonBackend includes error handling for common database connection and operation issues, raising ConnectionError or psycopg2.Error as appropriate.
• Ensure your Neon database instance is active and accessible from the environment where your application is running.
• Refer to the Neon documentation for details on managing your database, connection pooling, and security best practices.

Custom Backend Implementation

The llm-accounting library is designed with a pluggable backend system, allowing you to integrate with any database or data storage solution by implementing the BaseBackend abstract class. This is particularly useful for integrating with existing infrastructure or custom data handling requirements.

Here's how you can implement your own custom backend, using the MockBackend as a simplified example:

1. Define your Backend Class: Create a new class that inherits from llm_accounting.backends.base.BaseBackend. You will need to implement all abstract methods defined in BaseBackend.

   # my_custom_backend.py
   from datetime import datetime
   from typing import Dict, List, Tuple, Any, Optional
   
   from llm_accounting.backends.base import BaseBackend, UsageEntry, UsageStats, APIRequest # Added APIRequest
   
   class MyCustomBackend(BaseBackend):
       def __init__(self):
           self.usage_storage = [] # Example: a list to store UsageEntry objects
           self.request_storage = [] # Example: a list to store APIRequest objects
           # Add storage for limits if needed
   
       def initialize(self) -> None:
           print("MyCustomBackend: Initializing connection/resources...")
           # Implement your database connection or resource setup here
   
       def insert_usage(self, entry: UsageEntry) -> None:
           print(f"MyCustomBackend: Inserting usage for model {entry.model}")
           self.usage_storage.append(entry)
           # Implement logic to save 'entry' to your database
   
       def insert_api_request(self, request: APIRequest) -> None: # New method from BaseBackend
           print(f"MyCustomBackend: Inserting API request for model {request.model_name}")
           self.request_storage.append(request)
           # Implement logic to save 'request' to your database
   
       # ... (implement other abstract methods like get_period_stats, get_model_stats, etc.) ...
       # ... (get_model_rankings, purge, tail, close, execute_query) ...
       # ... (get_usage_limits, insert_usage_limit, get_api_requests_for_quota) ...
       # ... (get_usage_costs, set_usage_limit, get_usage_limit, record_api_request (from dict)) ...
   
       def get_period_stats(self, start: datetime, end: datetime) -> UsageStats:
           # Dummy implementation
           return UsageStats()
   
       def get_model_stats(self, start: datetime, end: datetime) -> List[Tuple[str, UsageStats]]:
           # Dummy implementation
           return []
       
       def get_model_rankings(self, start: datetime, end: datetime) -> Dict[str, List[Tuple[str, Any]]]:
           # Dummy implementation
           return {}
   
       def purge(self) -> None:
           self.usage_storage = []
           self.request_storage = []
       
       def tail(self, n: int = 10) -> List[UsageEntry]:
           return self.usage_storage[-n:]
   
       def close(self) -> None:
           print("MyCustomBackend: Closing connection/resources...")
   
       def execute_query(self, query: str) -> List[Dict[str, Any]]: # Corrected return type
           print(f"MyCustomBackend: Executing custom query: {query}")
           return []
           
       # You would also need to implement other methods from BaseBackend like:
       # get_usage_limits, insert_usage_limit, get_api_requests_for_quota,
       # get_usage_costs, set_usage_limit, get_usage_limit, record_api_request (dict version)
   

2. Integrate with LLMAccounting: Once your custom backend is implemented, you can pass an instance of it to the LLMAccounting constructor:

   from llm_accounting import LLMAccounting
   # from my_custom_backend import MyCustomBackend # Import your custom backend
   
   # Instantiate your custom backend
   # custom_backend = MyCustomBackend()
   
   # Pass it to LLMAccounting
   # accounting_custom = LLMAccounting(backend=custom_backend)
   
   # Now, all accounting operations will use your custom backend
   

with accounting_custom:

accounting_custom.track_usage(model="custom_model", prompt_tokens=10, cost=0.001)

# stats = accounting_custom.get_period_stats(datetime.now(), datetime.now())

# ... and so on

By following this pattern, you can extend `llm-accounting` to work seamlessly with virtually any data storage solution, providing maximum flexibility for your application's needs.

## Projects Utilizing LLM Accounting

We will be adding examples of projects that utilize `llm-accounting` in the nearest future to demonstrate reference usage.

## Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

## License

This project is licensed under the MIT License - see the LICENSE file for details.