ragflow/rag/llm/rerank_model.py

#
#  Copyright 2024 The InfiniFlow Authors. All Rights Reserved.
#
#  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
import threading
from urllib.parse import urljoin

import requests
import httpx
from huggingface_hub import snapshot_download
import os
from abc import ABC
import numpy as np
from yarl import URL

from api import settings
from api.utils.file_utils import get_home_cache_dir
from rag.utils import num_tokens_from_string, truncate
import json


def sigmoid(x):
    return 1 / (1 + np.exp(-x))


class Base(ABC):
    def __init__(self, key, model_name):
        pass

    def similarity(self, query: str, texts: list):
        raise NotImplementedError("Please implement encode method!")

    def total_token_count(self, resp):
        try:
            return resp.usage.total_tokens
        except Exception:
            pass
        try:
            return resp["usage"]["total_tokens"]
        except Exception:
            pass
        return 0


class DefaultRerank(Base):
    _model = None
    _model_lock = threading.Lock()

    def __init__(self, key, model_name, **kwargs):
        """
        If you have trouble downloading HuggingFace models, -_^ this might help!!

        For Linux:
        export HF_ENDPOINT=https://hf-mirror.com

        For Windows:
        Good luck
        ^_-

        """
        if not settings.LIGHTEN and not DefaultRerank._model:
            import torch
            from FlagEmbedding import FlagReranker
            with DefaultRerank._model_lock:
                if not DefaultRerank._model:
                    try:
                        DefaultRerank._model = FlagReranker(
                            os.path.join(get_home_cache_dir(), re.sub(r"^[a-zA-Z0-9]+/", "", model_name)),
                            use_fp16=torch.cuda.is_available())
                    except Exception:
                        model_dir = snapshot_download(repo_id=model_name,
                                                      local_dir=os.path.join(get_home_cache_dir(),
                                                                             re.sub(r"^[a-zA-Z0-9]+/", "", model_name)),
                                                      local_dir_use_symlinks=False)
                        DefaultRerank._model = FlagReranker(model_dir, use_fp16=torch.cuda.is_available())
        self._model = DefaultRerank._model
        self._dynamic_batch_size = 8
        self._min_batch_size = 1

    def torch_empty_cache(self):
        try:
            import torch
            torch.cuda.empty_cache()
        except Exception as e:
            print(f"Error emptying cache: {e}")

    def _process_batch(self, pairs, max_batch_size=None):
        """template method for subclass call"""
        old_dynamic_batch_size = self._dynamic_batch_size
        if max_batch_size is not None:
            self._dynamic_batch_size = max_batch_size
        res = []
        i = 0
        while i < len(pairs):
            current_batch = self._dynamic_batch_size
            max_retries = 5
            retry_count = 0
            while retry_count < max_retries:
                try:
                    # call subclass implemented batch processing calculation
                    batch_scores = self._compute_batch_scores(pairs[i:i + current_batch])
                    res.extend(batch_scores)
                    i += current_batch
                    self._dynamic_batch_size = min(self._dynamic_batch_size * 2, 8)
                    break
                except RuntimeError as e:
                    if "CUDA out of memory" in str(e) and current_batch > self._min_batch_size:
                        current_batch = max(current_batch // 2, self._min_batch_size)
                        self.torch_empty_cache()
                        retry_count += 1
                    else:
                        raise
            if retry_count >= max_retries:
                raise RuntimeError("max retry times, still cannot process batch, please check your GPU memory")
            self.torch_empty_cache()

        self._dynamic_batch_size = old_dynamic_batch_size
        return np.array(res)

    def _compute_batch_scores(self, batch_pairs, max_length=None):
        if max_length is None:
            scores = self._model.compute_score(batch_pairs)
        else:
            scores = self._model.compute_score(batch_pairs, max_length=max_length)
        scores = sigmoid(np.array(scores)).tolist()
        return scores

    def similarity(self, query: str, texts: list):
        pairs = [(query, truncate(t, 2048)) for t in texts]
        token_count = 0
        for _, t in pairs:
            token_count += num_tokens_from_string(t)
        batch_size = 4096
        res = self._process_batch(pairs, max_batch_size=batch_size)
        return np.array(res), token_count


class JinaRerank(Base):
    def __init__(self, key, model_name="jina-reranker-v2-base-multilingual",
                 base_url="https://api.jina.ai/v1/rerank"):
        self.base_url = "https://api.jina.ai/v1/rerank"
        self.headers = {
            "Content-Type": "application/json",
            "Authorization": f"Bearer {key}"
        }
        self.model_name = model_name

    def similarity(self, query: str, texts: list):
        texts = [truncate(t, 8196) for t in texts]
        data = {
            "model": self.model_name,
            "query": query,
            "documents": texts,
            "top_n": len(texts)
        }
        res = requests.post(self.base_url, headers=self.headers, json=data).json()
        rank = np.zeros(len(texts), dtype=float)
        for d in res["results"]:
            rank[d["index"]] = d["relevance_score"]
        return rank, self.total_token_count(res)


class YoudaoRerank(DefaultRerank):
    _model = None
    _model_lock = threading.Lock()

    def __init__(self, key=None, model_name="maidalun1020/bce-reranker-base_v1", **kwargs):
        if not settings.LIGHTEN and not YoudaoRerank._model:
            from BCEmbedding import RerankerModel
            with YoudaoRerank._model_lock:
                if not YoudaoRerank._model:
                    try:
                        YoudaoRerank._model = RerankerModel(model_name_or_path=os.path.join(
                            get_home_cache_dir(),
                            re.sub(r"^[a-zA-Z0-9]+/", "", model_name)))
                    except Exception:
                        YoudaoRerank._model = RerankerModel(
                            model_name_or_path=model_name.replace(
                                "maidalun1020", "InfiniFlow"))

        self._model = YoudaoRerank._model

    def similarity(self, query: str, texts: list):
        pairs = [(query, truncate(t, self._model.max_length)) for t in texts]
        token_count = 0
        for _, t in pairs:
            token_count += num_tokens_from_string(t)
        batch_size = 8
        res = self._process_batch(pairs, max_batch_size=batch_size)
        return np.array(res), token_count


class XInferenceRerank(Base):
    def __init__(self, key="xxxxxxx", model_name="", base_url=""):
        if base_url.find("/v1") == -1:
            base_url = urljoin(base_url, "/v1/rerank")
        if base_url.find("/rerank") == -1:
            base_url = urljoin(base_url, "/v1/rerank")
        self.model_name = model_name
        self.base_url = base_url
        self.headers = {
            "Content-Type": "application/json",
            "accept": "application/json",
            "Authorization": f"Bearer {key}"
        }

    def similarity(self, query: str, texts: list):
        if len(texts) == 0:
            return np.array([]), 0
        pairs = [(query, truncate(t, 4096)) for t in texts]
        token_count = 0
        for _, t in pairs:
            token_count += num_tokens_from_string(t)
        data = {
            "model": self.model_name,
            "query": query,
            "return_documents": "true",
            "return_len": "true",
            "documents": texts
        }
        res = requests.post(self.base_url, headers=self.headers, json=data).json()
        rank = np.zeros(len(texts), dtype=float)
        for d in res["results"]:
            rank[d["index"]] = d["relevance_score"]
        return rank, token_count


class LocalAIRerank(Base):
    def __init__(self, key, model_name, base_url):
        if base_url.find("/rerank") == -1:
            self.base_url = urljoin(base_url, "/rerank")
        else:
            self.base_url = base_url
        self.headers = {
            "Content-Type": "application/json",
            "Authorization": f"Bearer {key}"
        }
        self.model_name = model_name.split("___")[0]

    def similarity(self, query: str, texts: list):
        # noway to config Ragflow , use fix setting
        texts = [truncate(t, 500) for t in texts]
        data = {
            "model": self.model_name,
            "query": query,
            "documents": texts,
            "top_n": len(texts),
        }
        token_count = 0
        for t in texts:
            token_count += num_tokens_from_string(t)
        res = requests.post(self.base_url, headers=self.headers, json=data).json()
        rank = np.zeros(len(texts), dtype=float)
        if 'results' not in res:
            raise ValueError("response not contains results\n" + str(res))
        for d in res["results"]:
            rank[d["index"]] = d["relevance_score"]

        # Normalize the rank values to the range 0 to 1
        min_rank = np.min(rank)
        max_rank = np.max(rank)

        # Avoid division by zero if all ranks are identical
        if max_rank - min_rank != 0:
            rank = (rank - min_rank) / (max_rank - min_rank)
        else:
            rank = np.zeros_like(rank)

        return rank, token_count


class NvidiaRerank(Base):
    def __init__(
            self, key, model_name, base_url="https://ai.api.nvidia.com/v1/retrieval/nvidia/"
    ):
        if not base_url:
            base_url = "https://ai.api.nvidia.com/v1/retrieval/nvidia/"
        self.model_name = model_name

        if self.model_name == "nvidia/nv-rerankqa-mistral-4b-v3":
            self.base_url = os.path.join(
                base_url, "nv-rerankqa-mistral-4b-v3", "reranking"
            )

        if self.model_name == "nvidia/rerank-qa-mistral-4b":
            self.base_url = os.path.join(base_url, "reranking")
            self.model_name = "nv-rerank-qa-mistral-4b:1"

        self.headers = {
            "accept": "application/json",
            "Content-Type": "application/json",
            "Authorization": f"Bearer {key}",
        }

    def similarity(self, query: str, texts: list):
        token_count = num_tokens_from_string(query) + sum(
            [num_tokens_from_string(t) for t in texts]
        )
        data = {
            "model": self.model_name,
            "query": {"text": query},
            "passages": [{"text": text} for text in texts],
            "truncate": "END",
            "top_n": len(texts),
        }
        res = requests.post(self.base_url, headers=self.headers, json=data).json()
        rank = np.zeros(len(texts), dtype=float)
        for d in res["rankings"]:
            rank[d["index"]] = d["logit"]
        return rank, token_count


class LmStudioRerank(Base):
    def __init__(self, key, model_name, base_url):
        pass

    def similarity(self, query: str, texts: list):
        raise NotImplementedError("The LmStudioRerank has not been implement")


class OpenAI_APIRerank(Base):
    def __init__(self, key, model_name, base_url):
        if base_url.find("/rerank") == -1:
            self.base_url = urljoin(base_url, "/rerank")
        else:
            self.base_url = base_url
        self.headers = {
            "Content-Type": "application/json",
            "Authorization": f"Bearer {key}"
        }
        self.model_name = model_name.split("___")[0]

    def similarity(self, query: str, texts: list):
        # noway to config Ragflow , use fix setting
        texts = [truncate(t, 500) for t in texts]
        data = {
            "model": self.model_name,
            "query": query,
            "documents": texts,
            "top_n": len(texts),
        }
        token_count = 0
        for t in texts:
            token_count += num_tokens_from_string(t)
        res = requests.post(self.base_url, headers=self.headers, json=data).json()
        rank = np.zeros(len(texts), dtype=float)
        if 'results' not in res:
            raise ValueError("response not contains results\n" + str(res))
        for d in res["results"]:
            rank[d["index"]] = d["relevance_score"]

        # Normalize the rank values to the range 0 to 1
        min_rank = np.min(rank)
        max_rank = np.max(rank)

        # Avoid division by zero if all ranks are identical
        if max_rank - min_rank != 0:
            rank = (rank - min_rank) / (max_rank - min_rank)
        else:
            rank = np.zeros_like(rank)

        return rank, token_count


class CoHereRerank(Base):
    def __init__(self, key, model_name, base_url=None):
        from cohere import Client

        self.client = Client(api_key=key)
        self.model_name = model_name

    def similarity(self, query: str, texts: list):
        token_count = num_tokens_from_string(query) + sum(
            [num_tokens_from_string(t) for t in texts]
        )
        res = self.client.rerank(
            model=self.model_name,
            query=query,
            documents=texts,
            top_n=len(texts),
            return_documents=False,
        )
        rank = np.zeros(len(texts), dtype=float)
        for d in res.results:
            rank[d.index] = d.relevance_score
        return rank, token_count


class TogetherAIRerank(Base):
    def __init__(self, key, model_name, base_url):
        pass

    def similarity(self, query: str, texts: list):
        raise NotImplementedError("The api has not been implement")


class SILICONFLOWRerank(Base):
    def __init__(
            self, key, model_name, base_url="https://api.siliconflow.cn/v1/rerank"
    ):
        if not base_url:
            base_url = "https://api.siliconflow.cn/v1/rerank"
        self.model_name = model_name
        self.base_url = base_url
        self.headers = {
            "accept": "application/json",
            "content-type": "application/json",
            "authorization": f"Bearer {key}",
        }

    def similarity(self, query: str, texts: list):
        payload = {
            "model": self.model_name,
            "query": query,
            "documents": texts,
            "top_n": len(texts),
            "return_documents": False,
            "max_chunks_per_doc": 1024,
            "overlap_tokens": 80,
        }
        response = requests.post(
            self.base_url, json=payload, headers=self.headers
        ).json()
        rank = np.zeros(len(texts), dtype=float)
        if "results" not in response:
            return rank, 0

        for d in response["results"]:
            rank[d["index"]] = d["relevance_score"]
        return (
            rank,
            response["meta"]["tokens"]["input_tokens"] + response["meta"]["tokens"]["output_tokens"],
        )


class BaiduYiyanRerank(Base):
    def __init__(self, key, model_name, base_url=None):
        from qianfan.resources import Reranker

        key = json.loads(key)
        ak = key.get("yiyan_ak", "")
        sk = key.get("yiyan_sk", "")
        self.client = Reranker(ak=ak, sk=sk)
        self.model_name = model_name

    def similarity(self, query: str, texts: list):
        res = self.client.do(
            model=self.model_name,
            query=query,
            documents=texts,
            top_n=len(texts),
        ).body
        rank = np.zeros(len(texts), dtype=float)
        for d in res["results"]:
            rank[d["index"]] = d["relevance_score"]
        return rank, self.total_token_count(res)


class VoyageRerank(Base):
    def __init__(self, key, model_name, base_url=None):
        import voyageai

        self.client = voyageai.Client(api_key=key)
        self.model_name = model_name

    def similarity(self, query: str, texts: list):
        rank = np.zeros(len(texts), dtype=float)
        if not texts:
            return rank, 0
        res = self.client.rerank(
            query=query, documents=texts, model=self.model_name, top_k=len(texts)
        )
        for r in res.results:
            rank[r.index] = r.relevance_score
        return rank, res.total_tokens


class QWenRerank(Base):
    def __init__(self, key, model_name='gte-rerank', base_url=None, **kwargs):
        import dashscope
        self.api_key = key
        self.model_name = dashscope.TextReRank.Models.gte_rerank if model_name is None else model_name

    def similarity(self, query: str, texts: list):
        import dashscope
        from http import HTTPStatus
        resp = dashscope.TextReRank.call(
            api_key=self.api_key,
            model=self.model_name,
            query=query,
            documents=texts,
            top_n=len(texts),
            return_documents=False
        )
        rank = np.zeros(len(texts), dtype=float)
        if resp.status_code == HTTPStatus.OK:
            for r in resp.output.results:
                rank[r.index] = r.relevance_score
            return rank, resp.usage.total_tokens
        else:
            raise ValueError(f"Error calling QWenRerank model {self.model_name}: {resp.status_code} - {resp.text}")


class HuggingfaceRerank(DefaultRerank):
    @staticmethod
    def post(query: str, texts: list, url="127.0.0.1"):
        exc = None
        scores = [0 for _ in range(len(texts))]
        batch_size = 8
        for i in range(0, len(texts), batch_size):
            try:
                res = requests.post(f"http://{url}/rerank", headers={"Content-Type": "application/json"},
                                    json={"query": query, "texts": texts[i: i + batch_size],
                                          "raw_scores": False, "truncate": True})
                for o in res.json():
                    scores[o["index"] + i] = o["score"]
            except Exception as e:
                exc = e

        if exc:
            raise exc
        return np.array(scores)

    def __init__(self, key, model_name="BAAI/bge-reranker-v2-m3", base_url="http://127.0.0.1"):
        self.model_name = model_name
        self.base_url = base_url

    def similarity(self, query: str, texts: list) -> tuple[np.ndarray, int]:
        if not texts:
            return np.array([]), 0
        token_count = 0
        for t in texts:
            token_count += num_tokens_from_string(t)
        return HuggingfaceRerank.post(query, texts, self.base_url), token_count


class GPUStackRerank(Base):
    def __init__(
            self, key, model_name, base_url
    ):
        if not base_url:
            raise ValueError("url cannot be None")

        self.model_name = model_name
        self.base_url = str(URL(base_url) / "v1" / "rerank")
        self.headers = {
            "accept": "application/json",
            "content-type": "application/json",
            "authorization": f"Bearer {key}",
        }

    def similarity(self, query: str, texts: list):
        payload = {
            "model": self.model_name,
            "query": query,
            "documents": texts,
            "top_n": len(texts),
        }

        try:
            response = requests.post(
                self.base_url, json=payload, headers=self.headers
            )
            response.raise_for_status()
            response_json = response.json()

            rank = np.zeros(len(texts), dtype=float)
            if "results" not in response_json:
                return rank, 0

            token_count = 0
            for t in texts:
                token_count += num_tokens_from_string(t)

            for result in response_json["results"]:
                rank[result["index"]] = result["relevance_score"]

            return (
                rank,
                token_count,
            )

        except httpx.HTTPStatusError as e:
            raise ValueError(
                f"Error calling GPUStackRerank model {self.model_name}: {e.response.status_code} - {e.response.text}")