"""Generate an OpenSearch request with opensearch-agentic-dsl on a CUDA GPU.""" import json import os from pathlib import Path import torch from peft import PeftModel from transformers import AutoModelForImageTextToText, AutoTokenizer BASE_MODEL = "mistralai/Ministral-3-8B-Instruct-2512-BF16" BASE_REVISION = "06cc81bfd6e45321d8fc8f816576c5b6ac67ec22" ASSET_DIR = Path(__file__).resolve().parent ADAPTER_ID = os.environ.get("ADAPTER_ID", str(ASSET_DIR)) SYSTEM_PROMPT = (ASSET_DIR / "prompts/system.txt").read_text(encoding="utf-8").strip() USER_PROMPT_TEMPLATE = (ASSET_DIR / "prompts/user.txt").read_text(encoding="utf-8").strip() context = { "required_filters": [{"term": {"availability": "active"}}], "ui_max_price": 8000, "size": 24, "track_total_hits": 10000, "sort_mode": "recommended", "persona": {"id": "anonymous", "version": 1, "mode": "unprofiled"}, } question = ( "Shopper request: dress watch under 5000\n" "Trusted planner context: " + json.dumps(context, ensure_ascii=False, separators=(",", ":")) + "\nDerive price and gender affinity in the generated DSL. " "Apply persona only as instructed. Follow sort_mode." ) if len(question) > 1000: raise ValueError("Native planner question exceeds 1,000 characters") index_mapping = { "_doc": { "dynamic": "false", "properties": { "title": {"type": "text"}, "brand": {"type": "text"}, "canonical_text": {"type": "text"}, "description": {"type": "text"}, "availability": {"type": "keyword"}, "category": {"type": "keyword"}, "price": {"type": "integer"}, "quality_score": {"type": "rank_feature"}, "freshness_score": {"type": "rank_feature"}, "seller_score": {"type": "rank_feature"}, }, } } query_fields = list(index_mapping["_doc"]["properties"]) def native_string_parameter(value): compact = json.dumps(value, ensure_ascii=False, separators=(",", ":")) return json.dumps(compact, ensure_ascii=False) user_prompt = USER_PROMPT_TEMPLATE for placeholder, value in { "${parameters.question}": question, "${parameters.index_mapping:-}": native_string_parameter(index_mapping), "${parameters.query_fields:-}": native_string_parameter(query_fields), }.items(): user_prompt = user_prompt.replace(placeholder, value) messages = [ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": user_prompt}, ] tokenizer = AutoTokenizer.from_pretrained( BASE_MODEL, revision=BASE_REVISION, trust_remote_code=False ) tokenizer.pad_token = tokenizer.eos_token base_model = AutoModelForImageTextToText.from_pretrained( BASE_MODEL, revision=BASE_REVISION, dtype=torch.bfloat16, device_map={"": 0}, trust_remote_code=False, ) model = PeftModel.from_pretrained(base_model, ADAPTER_ID) model.eval() input_ids = tokenizer.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_dict=False ) inputs = { "input_ids": torch.tensor([input_ids], device=model.device), "attention_mask": torch.ones((1, len(input_ids)), dtype=torch.long, device=model.device), } with torch.inference_mode(): generated = model.generate( **inputs, do_sample=False, max_new_tokens=1024, eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.pad_token_id, ) completion = tokenizer.decode( generated[0, len(input_ids):], skip_special_tokens=True ) print(completion)