建立一个能搜索正确数据源并准确回答问题的ChatGPT

　　在这个例子中，我们将要求GPT在用户提出问题时挑选合适的数据集进行搜索，然后回答用户的问题。
　　，否则你可以通过使用Embedbase Cloud而不是自己运行它来简化这个例子。
　　如果是这样，你可以跳到种子数据集部分。安装
　　在虚拟环境中安装所需的依赖项：virtualenv env source env/bin/activate pip install embedbase pgvector psycopg2 openai 启动Postgres作为一个Embedbase数据库
　　为Embedbase数据库运行一个Postgres实例。docker run -d --name pgvector -p 8080:8080 -p 5432:5432   -e POSTGRES_DB=embedbase -e POSTGRES_PASSWORD=localdb   -v data:/var/lib/postgresql/data ankane/pgvector 启动嵌入基地
　　创建一个新文件main.py，代码如下：import os from embedbase import get_app from embedbase.database.postgres_db import Postgres from embedbase.embedding.openai import OpenAI import uvicorn OPENAI_API_KEY = os.environ.get(＂OPENAI_API_KEY＂) app = (       get_app()       .use_embedder(OpenAI(OPENAI_API_KEY))       .use_db(Postgres())       .run()   ) if __name__ == ＂__main__＂:     uvicorn.run(＂main:app＂, reload=True)
　　用以下命令启动Embedbase应用程序：python3 main.py 种子数据集
　　我们需要在Embedbase中添加一些数据来询问ChatGPT。import json import requests import fire   # Set the Embedbase API URL EMBEDBASE_API_URL = ＂http://localhost:8000＂ # if using embedbase cloud, add your api key to the headers # EMBEDBASE_API_KEY = ＂＂   def seed_dataset():     animals = {         ＂lion＂: {＂weight＂: 190, ＂height＂: 1.2, ＂speed＂: 80},         ＂elephant＂: {＂weight＂: 5000, ＂height＂: 3.2, ＂speed＂: 40},         ＂giraffe＂: {＂weight＂: 800, ＂height＂: 5.5, ＂speed＂: 60},         ＂zebra＂: {＂weight＂: 350, ＂height＂: 1.5, ＂speed＂: 60},         ＂rhinoceros＂: {＂weight＂: 2300, ＂height＂: 1.8, ＂speed＂: 45},         ＂crocodile＂: {＂weight＂: 1000, ＂height＂: 4.5, ＂speed＂: 20},         ＂hippopotamus＂: {＂weight＂: 1500, ＂height＂: 1.5, ＂speed＂: 30},         ＂cheetah＂: {＂weight＂: 60, ＂height＂: 0.8, ＂speed＂: 110},         ＂kangaroo＂: {＂weight＂: 80, ＂height＂: 1.5, ＂speed＂: 56},         ＂penguin＂: {＂weight＂: 30, ＂height＂: 1.1, ＂speed＂: 10},     }     cars = [         {＂make＂: ＂Toyota＂, ＂model＂: ＂Camry＂, ＂year＂: 2022},         {＂make＂: ＂Honda＂, ＂model＂: ＂Civic＂, ＂year＂: 2021},         {＂make＂: ＂Ford＂, ＂model＂: ＂F-150＂, ＂year＂: 2023},         {＂make＂: ＂Tesla＂, ＂model＂: ＂Model S＂, ＂year＂: 2022},         {＂make＂: ＂Chevrolet＂, ＂model＂: ＂Corvette＂, ＂year＂: 2021},         {＂make＂: ＂Jeep＂, ＂model＂: ＂Wrangler＂, ＂year＂: 2022},         {＂make＂: ＂BMW＂, ＂model＂: ＂X5＂, ＂year＂: 2023},         {＂make＂: ＂Mercedes-Benz＂, ＂model＂: ＂S-Class＂, ＂year＂: 2022},         {＂make＂: ＂Audi＂, ＂model＂: ＂A4＂, ＂year＂: 2021},         {＂make＂: ＂Lamborghini＂, ＂model＂: ＂Aventador＂, ＂year＂: 2022},     ]       # clear the dataset just in case it already exists     requests.get(f＂{EMBEDBASE_API_URL}/v1/animals/clear＂,         # if using embedbase cloud, add your api key to the headers         # headers={         #     ＂Authorization＂: ＂Bearer ＂ + EMBEDBASE_API_KEY,         # },     )     requests.get(f＂{EMBEDBASE_API_URL}/v1/cars/clear＂,         # if using embedbase cloud, add your api key to the headers         # headers={         #     ＂Authorization＂: ＂Bearer ＂ + EMBEDBASE_API_KEY,         # },     )       # seed the animals dataset     requests.post(         f＂{EMBEDBASE_API_URL}/v1/animals＂,         json={＂documents＂: [{＂data＂: json.dumps(animal)} for animal in animals]},         # if using embedbase cloud, add your api key to the headers         # headers={         #     ＂Authorization＂: ＂Bearer ＂ + EMBEDBASE_API_KEY,         # },     )       # seed the cars dataset     requests.post(         f＂{EMBEDBASE_API_URL}/v1/cars＂,         json={＂documents＂: [{＂data＂: json.dumps(car)} for car in cars]},         # if using embedbase cloud, add your api key to the headers         # headers={         #     ＂Authorization＂: ＂Bearer ＂ + EMBEDBASE_API_KEY,         # },     )   if __name__ == ＂__main__＂:     fire.Fire({         ＂seed＂: seed_dataset,     })python3 ask.py seed 搜索
　　我们现在将创建我们应用程序的主要逻辑。我们将要求GPT在用户提出问题时挑选正确的数据集进行搜索。
　　这个过程将按原样进行：
　　1. 用户提出一个问题
　　2. GPT将查询`/datasets`以获得数据集的列表
　　3. GPT将用所选择的数据集和问题查询`/search`。
　　4. GPT将返回结果import re import os import json import requests import openai import fire   # Set the Embedbase API URL EMBEDBASE_API_URL = ＂http://localhost:8000＂     def get_datasets():     response = requests.get(         f＂{EMBEDBASE_API_URL}/v1/datasets＂,         # if using embedbase cloud, add your api key to the headers         # headers={         #     ＂Authorization＂: ＂Bearer ＂ + EMBEDBASE_API_KEY,         # },     )     return [e[＂dataset_id＂] for e in response.json()[＂datasets＂]]     def search_dataset(dataset_id, query):     payload = {＂query＂: query, ＂top_k＂: 3}     response = requests.post(         f＂{EMBEDBASE_API_URL}/v1/{dataset_id}/search＂, json=payload,         # if using embedbase cloud, add your api key to the headers         # headers={         #     ＂Authorization＂: ＂Bearer ＂ + EMBEDBASE_API_KEY,         # },     )     return [e[＂data＂] for e in response.json()[＂similarities＂]]
　　上述代码将被用来查询Embedbase API。# ... def ask_question(question, openai_model: str = ＂gpt-3.5-turbo＂):     datasets = get_datasets()       # Prompt for GPT     prompt = f＂Given the following datasets: ＂     for dataset in datasets:         prompt += f＂- {dataset} ＂     prompt += f＂ Choose the best dataset to search and answer the following question: {question}＂       # Call GPT     response = openai.ChatCompletion.create(         model=openai_model,         messages=[             {                 ＂role＂: ＂system＂,                 ＂content＂: ＂You are a helpful assistant that select a dataset to search for a given question.＂                 ＂You always say ONLY the dataset name, nothing else. You are given a list of datasets and a question. ＂                 ＂For example, if the list of datasets is - plants - animals - cars - fruits - vegetables ＂                 ＂and the question is: What is the fastest animal?, you would say: [animals]＂,             },             {＂role＂: ＂user＂, ＂content＂: prompt},         ],     )       chosen_dataset = response.choices[0].message.content.strip()     print(f＂GPT chose the dataset: {chosen_dataset}＂)       # extract the dataset name from the output of GPT     # eg [animals] -> animals     chosen_dataset = re.sub(r＂[|]＂, ＂＂, chosen_dataset)     search_results = search_dataset(chosen_dataset, question)       # Call GPT again to answer the question based on the search results     prompt = (         f＂Based on the following search results, answer the question: ＂{question}＂ ＂     )     for result in search_results:         prompt += f＂- {result} ＂       response = openai.ChatCompletion.create(         model=openai_model,         messages=[             {                 ＂role＂: ＂system＂,                 ＂content＂: ＂You are a helpful assistant that answers questions based on the provided search results.＂,             },             {＂role＂: ＂user＂, ＂content＂: prompt},         ],     )       answer = response.choices[0].message.content.strip()       return answer
　　上述代码将被用来调用GPT来提问。现在添加一些小的逻辑，当用户提出问题时调用上述函数。 def main(openai_key: str = None, openai_model: str = ＂gpt-3.5-turbo＂):     openai.api_key = openai_key or os.environ.get(＂OPENAI_API_KEY＂)     question = input(＂Ask a question: ＂)     answer = ask_question(question, openai_model)     print(f＂Answer: {answer}＂) if __name__ == ＂__main__＂:     fire.Fire({               ＂ask＂: main,               ＂seed＂: seed_dataset,           })
　　现在你可以运行应用程序并提出问题。python3 ask.py ask --openai_key    # feel free to add ＂--openai_model gpt-4＂ if you have access to it