Eval AI Library
Based on firstlinesoftware/eval-ai-library. This is an independently maintained version with additional features and PyPI distribution.
Comprehensive AI model evaluation framework for RAG systems and AI agents. Supports 35+ evaluation metrics, 30+ LLM providers (OpenAI, Anthropic, Google, Azure, AWS Bedrock, Vertex AI, Mistral, Groq, xAI, DeepSeek, Cohere, OpenRouter, Together, Fireworks, Ollama, MLX, and more via LiteLLM), built-in test data generation from documents, an interactive web dashboard for visualization and analysis, and a first-class tracing subsystem with framework integrations for LangChain, LlamaIndex, CrewAI, AutoGen, Haystack, Semantic Kernel, Claude Agent SDK, smolagents, phidata, and OpenAI Assistants (plus an OpenTelemetry exporter). Implements advanced techniques including G-Eval probability-weighted scoring and Temperature-Controlled Verdict Aggregation via Generalized Power Mean.
Installation
pip install eval-ai-library
Full version with document parsing and OCR support:
pip install eval-ai-library[full]
Lite version (core evaluation only):
pip install eval-ai-library[lite]
Quick Start
from eval_lib import EvalAI
evaluator = EvalAI(model="gpt-4o")
result = evaluator.evaluate(
input="What is Python?",
actual_output="Python is a programming language.",
expected_output="Python is a high-level programming language.",
metrics=["answer_relevancy", "faithfulness"]
)
print(result.score)
Tracing
Capture agent runs (LLM calls, tool calls, steps) in your application and ship them to a collector for runtime evaluation. Three environment variables are enough to start:
TRACING_ENABLED=true
TRACING_URL=https://your-collector/api/traces/ingest
TRACING_PROJECT=my-agent
from eval_lib.tracing import tracer, trace_llm, trace_tool, SpanType
@trace_llm(name="answer")
async def answer(prompt: str) -> str:
...
trace_id = tracer.start_trace("chat")
result = await answer("hello")
tracer.add_trace_usage(input_tokens=120, output_tokens=30, cost_usd=0.0004) # per LLM call — accumulates
tracer.set_trace_metadata(model="gpt-4o", input="hello", output=result, # facts — overwrites
session_id="sess-1", customer="acme")
tracer.end_trace()
await tracer.aclose() # on shutdown: awaits in-flight deliveries, closes the HTTP pool
Two ways to attach data — don't mix them up. add_trace_usage(...) adds to running totals and is the right call once per LLM call. set_trace_metadata(...) declares a fact and overwrites; calling it per LLM call with that call's tokens keeps only the last call. Declared totals, when given, take precedence over accumulated ones.
Payload. Every trace is sent as {"project": ..., "trace": {...}} where the trace carries, alongside the span tree (spans, nested children):
| Field | What it is |
|---|---|
usage |
Accumulated counters: input_tokens, output_tokens, total_tokens, cached_tokens, reasoning_tokens, cost_usd, llm_calls, and source (accumulated / spans / declared) |
metadata |
Everything passed to set_trace_metadata, verbatim, as one object — map it straight onto your own metadata column |
started_at / ended_at |
ISO-8601 UTC (start_time / end_time epoch floats are kept for compatibility) |
model, input, output, session_id, user_id, cost_usd, input_tokens, … |
Promoted to top level for older consumers |
Configuration.
| Variable | Default | Purpose |
|---|---|---|
TRACING_ENABLED |
false |
Master switch |
TRACING_URL |
— | Collector endpoint (POST) |
TRACING_PROJECT |
default |
Project name sent with every payload |
TRACING_API_KEY |
— | Sent as Authorization: Bearer … |
TRACING_SINK |
http |
http / file (JSONL at TRACING_SINK_PATH, default traces.jsonl) / memory |
TRACING_STREAM |
false |
Ship each span as soon as it ends (partial_span), then the full trace at end_trace() — a crash mid-run still leaves spans on record |
TRACING_STRICT |
false |
Raise on delivery failure instead of logging a warning |
TRACING_MAX_RETRIES / TRACING_RETRY_BACKOFF |
2 / 0.5 |
Retries for transient failures (5xx, 408, 429, timeouts); 4xx fails fast |
TRACING_MAX_FIELD_LENGTH |
unlimited | Optional cap on captured span input/output; truncation is marked, never silent |
TRACING_REDACT |
true |
Redact credential-looking keys and token-shaped values in captured data |
Delivery is never silent: failures are logged under eval_lib.tracing, and tracer.stats exposes sent / failed / retried / dropped counters for health checks. Buffered traces are flushed at interpreter exit; in an async service call await tracer.aflush() (or aclose()) before shutdown so scheduled sends are not abandoned.
Framework integrations (eval_lib.tracing, each lazy-loaded so a slim install still imports): LangChain / LangGraph (EvalLibCallbackHandler), LlamaIndex (install_llamaindex_tracing, workflow and legacy APIs), CrewAI, AutoGen, Claude Agent SDK, OpenAI — Responses API and Chat Completions via trace_openai_client(OpenAI()) (function-call tool loops are paired across requests) plus the legacy Assistants API — Haystack, Semantic Kernel, smolagents, phidata / agno, and an OpenTelemetry SpanExporter (EvalLibSpanExporter). Manual tracing works from any code via the decorators or with tracer.trace("step"):. Note that a plain threading.Thread / thread-pool worker does not inherit the trace context — submit tracer.wrap(fn) instead of fn to carry it over, and end such a trace with tracer.end_trace(trace_id=...) if the worker finishes it.
Receiver. The bundled Flask receiver (eval_lib.connector.trace_routes) accepts both full traces and streamed partial_span payloads, stores metadata / usage unchanged, and flattens nested spans for the reliability metrics. Set TRACE_RECEIVER_ADMIN_KEY to protect project create/list when you are not supplying your own auth_verifier; a verifier returning {"projects": [...]} scopes the caller to those projects.
Documentation
Full documentation is available at library.eval-ai.com.
License
This project is licensed under the Apache License 2.0 - see the LICENSE file for details.
Citation
If you use this library in your research, please cite:
@software{eval_ai_library,
author = {Meshkov, Aleksandr},
title = {Eval AI Library: Comprehensive AI Model Evaluation Framework},
year = {2025},
url = {https://github.com/meshkovQA/Eval-ai-library.git}
}
References
This library implements techniques from:
@inproceedings{liu2023geval,
title={G-Eval: NLG Evaluation using GPT-4 with Better Human Alignment},
author={Liu, Yang and Iter, Dan and Xu, Yichong and Wang, Shuohang and Xu, Ruochen and Zhu, Chenguang},
booktitle={Proceedings of EMNLP},
year={2023}
}
Support
- Issues: GitHub Issues
- Documentation: library.eval-ai.com
1 maintainer
Metadata
1 maintainer
Release files for eval-ai-library 0.7.27
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| eval_ai_library-0.7.27.tar.gz | 485.5 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| eval_ai_library-0.7.27-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 883.2 kB
Release files / eval_ai_library-0.7.27.tar.gz
| Download URL | eval_ai_library-0.7.27.tar.gz |
|---|---|
| Size | 485.5 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
3a1458abe713499c4e876e16d889b70977fa3713d795ab0ed2ef9c4bf0e891c8
|
|
BLAKE2b-256 checksum How to use checksums |
d27b648a5f85a5aa745556a0b6083f08f498a14d5a6e726d7481bc40be0200f8
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.12.7
|
Release files / eval_ai_library-0.7.27-py3-none-any.whl
| Download URL | eval_ai_library-0.7.27-py3-none-any.whl |
|---|---|
| Size | 397.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
769725d5c961c11878e4a717323ec48c902b7cff1cf91722b07061b4ff79bb6e
|
|
BLAKE2b-256 checksum How to use checksums |
0aca3c05149b4bcb34aed49d0d416355809b1d475d61e672cd6d20b4d6d7856a
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.12.7
|
