Examples¶

Juicer usage examples. All examples call the vLLM OpenAI-compatible API directly — no extra framework needed.

Prerequisites¶

Start the Juicer model service:

vllm serve "$MODEL_ID" --port 8000 --tensor-parallel-size 2 --max-model-len 32768

Script	Purpose	Usage
`single_clean.py`	Minimal call: single-text cleaning	`python single_clean.py`
`batch_clean.py`	Batch cleaning: read JSONL → call API → write results	`python batch_clean.py input.jsonl output.jsonl`
`pii_redact.py`	PII redaction: email/phone placeholder replacement	`python pii_redact.py`

Juicer always outputs one of two contracts:

tagged_text (cleaning / filtering / redaction):

<status>KEEP</status><clean_text>processed text</clean_text>
<status>DROP</status><clean_text></clean_text>

json (tagging / scoring / classification):

{"helpfulness": 4, "correctness": 4, "coherence": 4, "complexity": 1, "verbosity": 2}

adapter.py (in the repo root) provides prompt construction and output parsing:

build_prompt_from_row(row) — build a prompt from a case row
parse_output(text, output_format) — parse tagged_text or json output
CLI: python adapter.py build-prompts ... / python adapter.py parse-outputs ...

See the Integration Code section in README.md for full code snippets.