aieng-eval-agents 0.3.1

Agentic AI Evaluation reference implementation from Vector Institute AI Engineering

aieng-eval-agents

Shared library for Vector Institute's Agentic AI Evaluation Bootcamp. Provides reusable components for building, running, and evaluating LLM agents with Google ADK and Langfuse.

What's included

Agent implementations

Module	Description
aieng.agent_evals.knowledge_qa	ReAct agent that answers questions using live web search. Includes evaluation against the DeepSearchQA benchmark with LLM-as-a-judge metrics (precision/recall/F1).
aieng.agent_evals.aml_investigation	Agent that investigates Anti-Money Laundering cases by querying a SQLite database of financial transactions via a read-only SQL tool.
aieng.agent_evals.report_generation	Agent that generates structured Excel reports from a relational database based on natural language queries.

Reusable tools (aieng.agent_evals.tools)

• search — Google Search with response grounding and citations
• web — HTML and PDF content fetching
• file — Download and search data files (CSV, XLSX, JSON)
• sql_database — Read-only SQL database access via ReadOnlySqlDatabase

Evaluation harness (aieng.agent_evals.evaluation)

Wrappers around Langfuse for running agent experiments:

• run_experiment — Run a dataset through an agent and score outputs
• run_experiment_with_trace_evals — Run experiments with trace-level evaluation
• run_trace_evaluations — Score existing Langfuse traces with LLM-based or heuristic graders

Utilities

• display — Rich-based terminal and Jupyter display helpers for metrics and agent responses
• progress — Progress tracking for batch evaluation runs
• configs — Pydantic-based configuration loading from .env
• langfuse — Langfuse client and trace utilities
• db_manager — Database connection management

Installation

pip install aieng-eval-agents

Requires Python 3.12+.

Source

Full reference implementations and documentation are in the eval-agents repository.