文献阅读与综述辅助

import json
import os
import re

from openai import OpenAI

client = OpenAI(base_url="https://www.tutujin.com/v1", api_key=os.environ["TUTUJIN_API_KEY"], max_retries=3)
FIELDS = ["research_question", "method", "data", "main_result", "limitations"]
SYSTEM = (
    "You fill one row of a literature table from the paper below and nothing else. For each field give a "
    "short value and the exact sentence of the paper it comes from. If the paper does not report a field, "
    'use {"value": "not reported", "quote": ""}. Reply with JSON only: an object with the keys '
    + ", ".join(FIELDS) + ', each {"value": "", "quote": ""}.'
)


def squash(s):
    return re.sub(r"\s+", " ", s or "").strip()


def extract(paper_id, text):
    resp = client.chat.completions.create(
        model="gemini-3.8-flash",
        messages=[{"role": "system", "content": SYSTEM}, {"role": "user", "content": text}],
        temperature=0,
    )
    raw = resp.choices[0].message.content
    row = json.loads(raw[raw.find("{"): raw.rfind("}") + 1])
    full = squash(text)
    for f in FIELDS:
        cell = row.get(f) if isinstance(row.get(f), dict) else {"value": "", "quote": ""}
        if cell.get("value") == "not reported":
            cell["check"] = "not_reported"
        else:
            # a value counts only if its sentence really is in the paper; everything else goes to a person
            quote = squash(cell.get("quote"))
            cell["check"] = "quote_found" if quote and quote in full else "by_hand"
        row[f] = cell
    return {"paper": paper_id, "model": resp.model, **{f: row[f] for f in FIELDS}}


def compare(rows, question):
    """Only rows whose quotes were all found are compared: the table is the source, not the model's memory."""
    table = [{"paper": r["paper"], **{f: r[f]["value"] for f in FIELDS}}
             for r in rows if all(r[f]["check"] != "by_hand" for f in FIELDS)]
    resp = client.chat.completions.create(
        model="claude-sonnet-5-5",
        messages=[
            {"role": "system", "content": "Compare the papers in the table and nothing else. Put the paper id in brackets, "
             "like [smith2024], after every statement. List agreements, contradictions and gaps. "
             "Do not add papers, facts or references that are not in the table."},
            {"role": "user", "content": "Question: " + question + "\n\nTable:\n" + json.dumps(table, ensure_ascii=False)},
        ],
    )
    return resp.choices[0].message.content


rows = []
for paper_id in ["smith2024", "li2025"]:  # full texts you are entitled to use, reference lists removed
    with open(f"papers/{paper_id}.txt", encoding="utf-8") as f:
        rows.append(extract(paper_id, f.read()))
print(json.dumps(rows, ensure_ascii=False, indent=2))  # a researcher checks every row against the paper
print(compare(rows, "Does remote tutoring improve test scores?"))
# Reading list: embed abstracts with text-embedding-3-large and search them as in the knowledge-base solution.