Command line

Installing the package adds a rag-titan command. python -m rag_titan runs the same tool with the same options; use it on Windows when security settings block the rag-titan.exe launcher, see Getting started.

Ask one question

rag-titan ask --files handbook.pdf --question "How many leave days do I get?"

Chat in a loop

rag-titan chat --files handbook.pdf policy.docx --rag-type adaptive

Type exit to quit.

See every step

rag-titan ask --files handbook.pdf --question "How many leave days do I get?" --trace

Prints every step the RAG type took (search results with their files and scores, evidence checks, rewrites) before the answer.

Evaluate and compare

rag-titan evaluate --files docs/*.md --questions questions.json \
    --rag-type simple adaptive --chunker word paragraph

Runs every RAG type with every chunker on your test questions and prints a report, then the questions that missed evidence or got no answer. Add --retrieval-only to score search alone (fast, no LLM). See Evaluating RAG.

Evaluate logged data

rag-titan evaluate --input data.json --output report.json --k 5 \
    --baseline last_report.json --fail-on-regression

Scores the logged outputs of any RAG system, stage by stage: chunking, embedding, retrieval, reranking, context, answers and end to end. It saves report.json and report.md and prints the overall score. With --baseline it lists the metrics that got worse, and --fail-on-regression then exits with code 1, for scheduled runs. No model is needed. See Evaluating every stage.

Write golden questions

rag-titan golden --questions questions.txt --files policy.md --output golden.json

Writes a first draft of your golden questions: one record per question (one per line in questions.txt, or a JSON list), with a suggested evidence sentence from your files. Every record starts with "validated": false: you review it, write the expected answer, then mark it validated. It never overwrites an existing file unless you add --force. See Build a golden set.

Options

Option	Used by	Meaning
--files	ask, chat, evaluate	PDF, DOCX, MD or TXT files (required, except with --input)
--rag-type	ask, chat, evaluate	simple (default) or adaptive; evaluate takes several
--log-level	ask, chat, evaluate	DEBUG, INFO or WARNING (default)
--question	ask	the question to answer (required)
--trace	ask	also print every step taken
--questions	evaluate	questions JSON file (required with --files)
--chunker	evaluate	one or more chunkers to compare
--chunk-size, --chunk-overlap	evaluate	used by every setup
--retrieval-only	evaluate	score search only, no LLM
--top-k	evaluate	chunks per search with --retrieval-only (default 5)
--input	evaluate	JSON data file to score, instead of --files and --questions
--output	evaluate	JSON report path with --input (default rag_report.json); a .md is saved next to it
--k	evaluate	results scored per query with --input (default 5)
--baseline	evaluate	an earlier JSON report: list what got worse
--fail-on-regression	evaluate	exit with code 1 when a metric got worse than the baseline
--questions	golden	questions, one per line (.txt) or a JSON list (required)
--files	golden	your documents, to suggest evidence quotes
--output	golden	where to write the template (default golden.json)
--force	golden	overwrite --output if it exists
--version	–	print the version

The CLI also reads your .env file, see Configuration.