CLI and result files

Kind: reference.

The executable is inference-axi. Run it through uv run from this checkout. Each command accepts --help. The parser rejects unknown flags before it calls external tools.

Command	Result
No arguments or status	Container state, readiness and free GPU memory
corpus --output <directory>	Eight images and a manifest with known text
build	Local Podman server image
start	Owned server container after a memory check
stop	Removal of the owned container
logs	Last 40 log lines
logs --full	All log lines
ocr <image> --output <file>	Raw model text and a JSON receipt
bench <corpus> --output <new-directory>	JSON results and a Markdown summary

ocr and bench accept a positive --timeout value in seconds. The default is 120 seconds per request. ocr accepts repeatable --expect phrases and an optional --url. bench accepts --concurrency values from 1, 2, 4 and 8. It requires this checkout's local server. It rejects an existing output directory.

The client converts each image to RGB PNG before it starts inference timing. It preserves class and coordinate tokens in the raw response. The raw output is not cleaned Markdown. PDF conversion is outside this demo.

Output channels and exit codes

Stdout contains TOON scalar fields. Stderr contains build diagnostics. The default OCR preview has at most 500 characters. The output file always contains the full text. ocr --full prints all text from a new request.

Exit code	Meaning
0	Command succeeded, including an idempotent stop
1	Runtime failure or unsuccessful OCR completion
2	Invalid command arguments
130	User interrupted the operation

Quality checks

The benchmark checks every expected phrase from the corpus manifest. It ignores case, punctuation and whitespace differences. It rejects empty output, missing phrases and incomplete generation. It counts HTTP errors and timeouts as failures.

ocr applies the same checks to each supplied --expect phrase. Without these phrases, it checks only nonempty text and normal generation completion. This weaker check does not establish text accuracy.

Result files

File	Contents
OCR output path	Raw model text, including unsuccessful output
OCR output path plus .json	Status, latency, input hash, output and error
results.json	Metadata, warmup and per-request sweep records
summary.md	Counts, throughput, latency statistics and metadata

A request has one terminal status. Statuses are success, error, timeout, incomplete and quality_failed. Only success contributes to successful pages per second.

Each sweep records input count, successes, failures, timeouts and elapsed seconds. Latency statistics contain count, minimum, median and maximum. The report separates all requests from successful requests. It retains every observed latency and raw output.

Metadata records input hashes, expected phrases and preparation time. It also records the model revision, serving configuration, image ID and installed packages. Hardware fields include GPU name, driver version and available memory. Source fields include the Git commit and working-tree state.

The runner writes results after each completed sweep. An aborted run retains completed sweeps and reports status: aborted. In-flight requests do not become completed observations.