Rag Evaluator

Python SDK for Agent AI Observability, Monitoring and Evaluation Framework. Includes features like a ragaai catalyst, python, agentic-ai.

AI-powered RAG (Retrieval-Augmented Generation) evaluation toolkit. Configure, benchmark, compare, and optimize your RAG pipelines from the command line. Track prompts, evaluations, fine-tuning experiments, costs, and usage — all with persistent local logging and full export capabilities.

Commands

Run rag-evaluator <command> [args] to use.

Command	Description
configure	Configure RAG evaluation settings and parameters
benchmark	Run benchmarks against your RAG pipeline
compare	Compare results across different RAG configurations
prompt	Log and manage prompt templates and variations
evaluate	Evaluate RAG output quality and relevance
fine-tune	Track fine-tuning experiments and parameters
analyze	Analyze evaluation results and identify patterns
cost	Track and log API/inference costs
usage	Monitor token usage and API call volumes
optimize	Log optimization strategies and results
test	Run test cases against RAG configurations
report	Generate evaluation reports
stats	Show summary statistics across all categories
export <fmt>	Export data in json, csv, or txt format
search <term>	Search across all logged entries
recent	Show recent activity from history log
status	Health check — version, data dir, disk usage
help	Show help and available commands
version	Show version (v2.0.0)

Each domain command (configure, benchmark, compare, etc.) works in two modes:

• Without arguments: displays the most recent 20 entries from that category
• With arguments: logs the input with a timestamp and saves to the category log file

Data Storage

All data is stored locally in ~/.local/share/rag-evaluator/:

• Each command creates its own log file (e.g., configure.log, benchmark.log)
• A unified history.log tracks all activity across commands
• Entries are stored in timestamp|value pipe-delimited format
• Export supports JSON, CSV, and plain text formats

Requirements

• Bash 4+ with set -euo pipefail strict mode
• Standard Unix utilities: date, wc, du, tail, grep, sed, cat
• No external dependencies or API keys required

When to Use

1. Evaluating RAG pipeline quality — log evaluation scores, compare retrieval strategies, and track improvements over time
2. Benchmarking different configurations — run benchmarks across embedding models, chunk sizes, or retrieval methods and compare results side by side
3. Tracking costs and usage — monitor API costs and token usage across experiments to stay within budget
4. Managing prompt engineering — log prompt variations, test them against your pipeline, and analyze which templates perform best
5. Generating reports for stakeholders — export evaluation data as JSON/CSV for dashboards, or generate text reports summarizing RAG performance

Examples

# Configure a new evaluation run
rag-evaluator configure "model=gpt-4 chunks=512 overlap=50 top_k=5"

# Run a benchmark and log results
rag-evaluator benchmark "latency=230ms recall@5=0.82 precision@5=0.71"

# Compare two retrieval strategies
rag-evaluator compare "bm25 vs dense: bm25 recall=0.78, dense recall=0.85"

# Track evaluation scores
rag-evaluator evaluate "faithfulness=0.91 relevance=0.87 coherence=0.93"

# Log API cost for a run
rag-evaluator cost "run-042: $0.23 (1.2k tokens input, 800 tokens output)"

# View summary statistics
rag-evaluator stats

# Export all data as CSV
rag-evaluator export csv

# Search for specific entries
rag-evaluator search "gpt-4"

# Check recent activity
rag-evaluator recent

# Health check
rag-evaluator status

Output

All commands output to stdout. Redirect to a file if needed:

rag-evaluator report "weekly summary" > report.txt
rag-evaluator export json  # saves to ~/.local/share/rag-evaluator/export.json

Configuration

Set DATA_DIR by modifying the script, or use the default: ~/.local/share/rag-evaluator/

---

Powered by BytesAgain | bytesagain.com | [e-mail]