rag/app/qa.py

#  Licensed under the Apache License, Version 2.0 (the "License");
#  you may not use this file except in compliance with the License.
#  You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
#  Unless required by applicable law or agreed to in writing, software
#  distributed under the License is distributed on an "AS IS" BASIS,
#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#  See the License for the specific language governing permissions and
#  limitations under the License.
#
import re
from copy import deepcopy
from io import BytesIO
from nltk import word_tokenize
from openpyxl import load_workbook
from rag.nlp import is_english, random_choices
from rag.nlp import huqie
from deepdoc.parser import ExcelParser


class Excel(ExcelParser):
    def __call__(self, fnm, binary=None, callback=None):
        if not binary:
            wb = load_workbook(fnm)
        else:
            wb = load_workbook(BytesIO(binary))
        total = 0
        for sheetname in wb.sheetnames:
            total += len(list(wb[sheetname].rows))

        res, fails = [], []
        for sheetname in wb.sheetnames:
            ws = wb[sheetname]
            rows = list(ws.rows)
            for i, r in enumerate(rows):
                q, a = "", ""
                for cell in r:
                    if not cell.value:
                        continue
                    if not q:
                        q = str(cell.value)
                    elif not a:
                        a = str(cell.value)
                    else:
                        break
                if q and a:
                    res.append((q, a))
                else:
                    fails.append(str(i + 1))
                if len(res) % 999 == 0:
                    callback(len(res) *
                             0.6 /
                             total, ("Extract Q&A: {}".format(len(res)) +
                                     (f"{len(fails)} failure, line: %s..." %
                                      (",".join(fails[:3])) if fails else "")))

        callback(0.6, ("Extract Q&A: {}. ".format(len(res)) + (
            f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))
        self.is_english = is_english(
            [rmPrefix(q) for q, _ in random_choices(res, k=30) if len(q) > 1])
        return res


def rmPrefix(txt):
    return re.sub(
        r"^(问题|答案|回答|user|assistant|Q|A|Question|Answer|问|答)[\t:： ]+", "", txt.strip(), flags=re.IGNORECASE)


def beAdoc(d, q, a, eng):
    qprefix = "Question: " if eng else "问题："
    aprefix = "Answer: " if eng else "回答："
    d["content_with_weight"] = "\t".join(
        [qprefix + rmPrefix(q), aprefix + rmPrefix(a)])
    d["content_ltks"] = huqie.qie(q)
    d["content_sm_ltks"] = huqie.qieqie(d["content_ltks"])
    return d


def chunk(filename, binary=None, lang="Chinese", callback=None, **kwargs):
    """
        Excel and csv(txt) format files are supported.
        If the file is in excel format, there should be 2 column question and answer without header.
        And question column is ahead of answer column.
        And it's O.K if it has multiple sheets as long as the columns are rightly composed.

        If it's in csv format, it should be UTF-8 encoded. Use TAB as delimiter to separate question and answer.

        All the deformed lines will be ignored.
        Every pair of Q&A will be treated as a chunk.
    """
    eng = lang.lower() == "english"
    res = []
    doc = {
        "docnm_kwd": filename,
        "title_tks": huqie.qie(re.sub(r"\.[a-zA-Z]+$", "", filename))
    }
    if re.search(r"\.xlsx?$", filename, re.IGNORECASE):
        callback(0.1, "Start to parse.")
        excel_parser = Excel()
        for q, a in excel_parser(filename, binary, callback):
            res.append(beAdoc(deepcopy(doc), q, a, eng))
        return res
    elif re.search(r"\.(txt|csv)$", filename, re.IGNORECASE):
        callback(0.1, "Start to parse.")
        txt = ""
        if binary:
            txt = binary.decode("utf-8")
        else:
            with open(filename, "r") as f:
                while True:
                    l = f.readline()
                    if not l:
                        break
                    txt += l
        lines = txt.split("\n")
        #is_english([rmPrefix(l) for l in lines[:100]])
        fails = []
        for i, line in enumerate(lines):
            arr = [l for l in line.split("\t") if len(l) > 1]
            if len(arr) != 2:
                fails.append(str(i))
                continue
            res.append(beAdoc(deepcopy(doc), arr[0], arr[1], eng))
            if len(res) % 999 == 0:
                callback(len(res) * 0.6 / len(lines), ("Extract Q&A: {}".format(len(res)) + (
                    f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))

        callback(0.6, ("Extract Q&A: {}".format(len(res)) + (
            f"{len(fails)} failure, line: %s..." % (",".join(fails[:3])) if fails else "")))

        return res

    raise NotImplementedError(
        "file type not supported yet(pptx, pdf supported)")


if __name__ == "__main__":
    import sys

    def dummy(a, b):
        pass
    chunk(sys.argv[1], callback=dummy)