模型适合处理「单行或单块的语义判断」,不适合做聚合计算。正确做法是:表格里加一列,逐行调用,再把结果写回。
每行数据独立处理,可以把「行 ID」一起发过去,返回时带回来做对齐:
{"id": "row-001", "title": "蓝牙耳机 降噪", "price": "199"}
需要跨行理解的内容(如同一订单的多条明细)再按业务主键聚合后再发。
import json, csv, time from openai import OpenAI client = OpenAI(api_key=KEY, base_url="https://hardysapi.com/v1") rows = list(csv.DictReader(open("in.csv", encoding="utf-8"))) done = {json.loads(l)["id"] for l in open("out.jsonl", encoding="utf-8")} with open("out.jsonl", "a", encoding="utf-8") as f: for r in rows: if r["id"] in done: continue for attempt in range(3): try: resp = client.chat.completions.create( model="glm-4.6", temperature=0, response_format={"type": "json_object"}, messages=[ {"role": "system", "content": SYS}, {"role": "user", "content": json.dumps(r, ensure_ascii=False)}, ]) f.write(json.dumps(json.loads(resp.choices[0].message.content), ensure_ascii=False) + "\n") f.flush() break except Exception as e: time.sleep(2 ** attempt) else: f.write(json.dumps({"id": r["id"], "error": "failed"}) + "\n")
json_object 模式 + 必填字段校验,缺字段视为失败;行数多时会超出上下文且难以对齐;逐行或按业务主键分块发送,返回结果带行 ID,才能可靠回写。
只发送必要列、使用轻量档模型、用 json_object 限制输出长度,并先跑 100 行样本估算总成本。