README.md

Flow Benchmark Tools

Create and run LLM benchmarks.

Installation

Just the library:

pip install flow-benchmark-tools:1.1.0

Library + Example benchmarks (see below):

pip install "flow-benchmark-tools[examples]:1.1.0"

Usage

• Create an agent by inheriting BenchmarkAgent and implementing the run_benchmark_case method.

• Create a Benchmark by compiling a list of BenchmarkCases. These can be read from a JSONL file.

• Associate agent and benchmark in a BenchmarkRun.

• Use a BenchmarkRunner to run your BenchmarkRun.

Running example RAG benchmarks

Two end-to-end benchmark examples are provided in the examples folder: a LangChain RAG application and an OpenAI Assistant agent.

To run the LangChain RAG benchmark:

python src/examples/langchain_rag_agent.py

To run the OpenAI Assistant benchmark:

python src/examples/openai_assistant_agent.py

The rag benchmark cases are defined in data/rag_benchmark.jsonl.

The two examples follow the typical usage pattern of the library:

• define an agent by implementing the BenchmarkAgent interface and overriding the run_benchmark_case method (you can also override the before and after methods, if needed),
• create a set of benchmark cases, typically as a JSONL file such as data/rag_benchmark.jsonl,
• use a BenchmarkRunner to run the benchmark.

Running example criteria benchmark

An application of a criteria benchmark is also provided in examples folder: a Criteria application that assesses the quality of pre-computed LLM outputs based on the criteria defined in each benchmark case.

To run the Criteria benchmark:

python src/examples/criteria_evaluation_agent.py

The criteria benchmark cases are defined in data/criteria_benchmark.jsonl.

This example follows a different application of the library:

• define an agent implementing the BenchmarkAgent interface. In this application, each case already has the output we want to evaluate, so we override the run_benchmark_case method to simply repackage each BenchmarkCase as BenchmarkCaseResponse.
• create a set of quality benchmark cases, typically as a JSONL file such as data/criteria_benchmark.jsonl. In this application, each case's "extra" dictionary includes a "criteria" string.
• use a custom CriteriaBenchmarkRunner which overrides the _execute_benchmark_case method, to run the benchmark using an evaluator that inherits from CriteriaEvaluator