nexa-gauge 0.1.6

A graph-based toolkit for evaluating LLM and RAG outputs with repeatable quality metrics and reporting.

nexa-gauge - Graph-Based Evaluation for LLM and RAG Systems

A cache-aware evaluation engine for measuring LLM and RAG output quality with repeatable metrics, cost estimates, and structured reports.

Read the Documentation  ·  Quickstart  ·  CLI Usage  ·  Report Bug  ·  Request Feature

---

What is nexa-gauge?

nexa-gauge is a Python package and command-line toolkit for evaluating generated outputs from LLM, RAG, and agentic systems. It replaces ad-hoc manual checks with a typed evaluation graph that can estimate cost, execute only the required nodes, reuse cached artifacts, and emit structured per-case reports.

It is designed for prompt iteration, benchmark runs, regression testing, release gates, and production evaluation workflows where teams need measurable quality and safety signals.

Core evaluation coverage includes:

• Relevance - measures whether generated claims answer the user question.
• Grounding - checks whether generated claims are supported by supplied context.
• Red team scoring - evaluates safety and risk behavior with configurable rubrics.
• GEval / LLM-as-a-judge - scores outputs against explicit criteria or evaluation steps.
• Reference metrics - computes overlap-based metrics against known reference answers.

---

Quickstart

Install the package from PyPI:

pip install nexa-gauge

Install optional Hugging Face dataset support:

pip install "nexa-gauge[huggingface]"

Set your provider key:

export OPENAI_API_KEY="<your-key>"

Estimate cost before running billable evaluation work:

nexagauge estimate eval --input sample.json --limit 10

Run the full evaluation graph and write per-case reports:

nexagauge run eval --input sample.json --limit 10 --output-dir ./report

---

Core Capabilities

• Graph-based evaluation pipeline - predictable node topology for scanning, chunking, claim extraction, metric execution, aggregation, and reporting.
• Estimate-first execution - preview uncached eligible cost before making LLM-backed calls.
• Cache-aware runs - avoid duplicate LLM spend and recomputation when inputs, prompts, and model routes are unchanged.
• Structured JSON reports - write per-case report files for CI, dashboards, notebooks, and downstream analytics.
• Per-node model routing - configure global models, node-specific models, fallback models, and temperatures.
• Scalable CLI execution - tune case-level workers, in-flight cases, and global LLM concurrency.
• Local and hosted datasets - evaluate JSON, JSONL, CSV, text files, and Hugging Face datasets.

---

Evaluation Pipeline

nexa-gauge runs evaluations through a deterministic node graph. Each target executes only its required upstream dependencies.

Category	Nodes	Purpose
Input and orchestration	scan, eval, report	Normalize records, aggregate metric branches, and project final reports.
Utility	chunk, claims, dedup, geval_steps	Prepare generated text, extract claims, remove duplicates, and resolve GEval steps.
Metrics	relevance, grounding, redteam, geval, reference	Score answer quality, evidence support, safety, rubric alignment, and reference overlap.

Typical execution paths:

grounding: scan -> chunk -> claims -> dedup -> grounding
relevance: scan -> chunk -> claims -> dedup -> relevance
geval:     scan -> geval_steps -> geval
eval:      full graph execution and aggregate metric summary

For a full architecture diagram, see docs/architecture.md.

---

CLI Usage

The CLI entry point is nexagauge.

nexagauge --help
nexagauge run --help
nexagauge estimate --help
nexagauge cache --help

Primary commands:

Command	Purpose
nexagauge estimate <target_node> --input <source>	Estimate uncached cost for a target branch before execution.
nexagauge run <target_node> --input <source>	Execute a target branch and optionally write reports.
nexagauge cache dir	Print the resolved cache root directory.
nexagauge cache delete	Inspect or clear cached node outputs.

Common examples:

# Estimate full evaluation cost for a dataset slice
nexagauge estimate eval --input sample.json --limit 100

# Run full evaluation and write JSON reports
nexagauge run eval --input sample.json --limit 100 --output-dir ./report

# Run only the grounding metric branch
nexagauge run grounding --input sample.json --limit 25

# Preview cache cleanup
nexagauge cache delete --dry-run

Common flags:

Area	Flags
Data selection	--input, --adapter, --start, --end, --limit
Model routing	--model, --llm-model, --llm-fallback
Caching	--force, --no-cache
Execution	--max-workers, --max-in-flight, --llm-concurrency, --continue-on-error
Debugging	--debug
Reports	--output-dir

See docs/cli-code-flow.md and the hosted CLI documentation for deeper usage details.

---

Input Formats

Use --input with local files or hosted datasets.

Source	Example	Notes
JSON	sample.json	Object or array of records.
JSONL	dataset.jsonl	One JSON object per line.
CSV	dataset.csv	Rows are loaded as dictionaries.
Text	generation.txt	Treated as a single generated output.
Hugging Face	hf://org/dataset	Requires pip install "nexa-gauge[huggingface]".

Canonical record fields include:

Field	Used by
generation	Required for all metric branches.
question	Relevance and some GEval configurations.
context	Grounding and context-aware GEval checks.
reference	Reference metrics and reference-aware GEval checks.
geval	Rubric-driven GEval metrics.
redteam	Custom safety and risk rubrics.

Common aliases such as response, answer, output, completion, query, prompt, ground_truth, and label are normalized during scanning.

---

Metrics

nexa-gauge combines deterministic metrics with LLM-as-a-judge evaluation.

Metric node	What it evaluates
relevance	Whether generated claims directly answer the question.
grounding	Whether generated claims are supported by the provided context.
redteam	Bias, toxicity, and custom risk behavior using rubrics.
geval	Criteria-based LLM judging with generated or provided evaluation steps.
reference	BLEU, METEOR, ROUGE-1, ROUGE-2, and ROUGE-L against reference answers.

GEval is split into two phases:

1. geval_steps resolves reusable evaluation steps from criteria or accepts provided steps.
2. geval scores each case against those resolved steps and selected input fields.

This design makes rubric-based evaluation repeatable and cache-friendly across datasets.

---

Caching and Cost Estimation

Cost control is a first-class part of the runtime.

# Preview uncached work before execution
nexagauge estimate eval --input sample.json --limit 50

# Reuse cache during normal runs
nexagauge run eval --input sample.json --limit 50 --output-dir ./report

# Ignore cache reads but still write fresh outputs
nexagauge run eval --input sample.json --limit 50 --force

# Disable cache reads and writes for debugging
nexagauge run eval --input sample.json --limit 50 --no-cache

The cache is deterministic and route-aware. Inputs, evaluation criteria, model routing, prompt versions, parser versions, and relevant upstream artifacts are included in cache keys so stale outputs are not reused across incompatible runs.

For run, cache location can be controlled with:

export NEXAGAUGE_CACHE_DIR="./.nexagauge-cache"

Inspect the active cache root:

nexagauge cache dir

Clear cached node outputs:

nexagauge cache delete --dry-run
nexagauge cache delete --yes

---

Configuration

For local development or repeatable runs, copy the environment template:

cp .env.example .env

Minimum configuration for OpenAI-backed runs:

OPENAI_API_KEY=<your-key>
LLM_MODEL=gpt-4o-mini

Supported per-node overrides follow this pattern:

LLM_CLAIMS_MODEL=openai/gpt-4o-mini
LLM_CLAIMS_FALLBACK_MODEL=openai/gpt-4o
LLM_GROUNDING_TEMPERATURE=0.0

Runtime overrides can also be passed through the CLI:

nexagauge run eval \
  --input sample.json \
  --llm-model openai/gpt-4o-mini \
  --llm-model grounding=openai/gpt-4o \
  --llm-fallback openai/gpt-4o

---

Development

Clone the repository and install it from source:

git clone https://github.com/harnexa/nexa-gauge.git
cd nexa-gauge
pip install -e .

Contributor workflow with uv:

uv sync
make lint
make test
make ci

Build distributions:

uv build

Expected artifacts:

dist/nexa_gauge-<version>-py3-none-any.whl
dist/nexa_gauge-<version>.tar.gz

Releases use release-please. Conventional Commit titles such as feat:, fix:, docs:, deps:, and chore: are recommended for cleaner generated release notes, but they are not required by CI.

---

Documentation

• Hosted documentation: harnexa.dev/nexa-gauge/docs/introduction
• Local getting started guide: docs/get-started.md
• Architecture: docs/architecture.md
• CLI code flow: docs/cli-code-flow.md
• Product summary: docs/PRODUCT_SUMMARY.md

---

Project Standards

• License: MIT
• Security policy: SECURITY.md
• Contributing guide: CONTRIBUTING.md
• Code of conduct: CODE_OF_CONDUCT.md

---

License

Distributed under the MIT License. See LICENSE for details.