使用搜索 API 和重新排序进行问答 | OpenAI Cookbook

搜索相关信息有时感觉像大海捞针，但不要绝望，GPT 实际上可以为我们完成很多这项工作。在本指南中，我们探索了一种使用各种 AI 技术增强现有搜索系统的方法，帮助我们筛选噪音。

为 GPT 检索信息的两种方法是

模仿人类浏览： GPT 触发搜索，评估结果，并在必要时修改搜索查询。它还可以跟进特定的搜索结果以形成思路链，就像人类用户一样。
使用嵌入检索： 计算您的内容和用户查询的嵌入，然后检索余弦相似度衡量的最相关内容。谷歌等搜索引擎大量使用这项技术。

这些方法都很有前景，但各有缺点：第一种方法由于其迭代性质可能很慢，第二种方法需要提前嵌入您的整个知识库，不断嵌入新内容并维护向量数据库。

通过结合这些方法，并从重新排序方法中汲取灵感，我们找到了一种介于两者之间的方法。这种方法可以在任何现有的搜索系统之上实现，例如 Slack 搜索 API，或具有私有数据的内部 ElasticSearch 实例。以下是它的工作原理

search_augmented_by_query_generation_and_embeddings_reranking.png

步骤 1：搜索

用户提出问题。
GPT 生成潜在查询列表。
并行执行搜索查询。

步骤 2：重新排序

每个结果的嵌入用于计算与生成的理想假设答案的用户问题之间的语义相似度。
结果根据此相似性指标进行排名和筛选。

步骤 3：回答

给定顶部搜索结果，模型生成用户问题的答案，包括参考文献和链接。

这种混合方法提供相对较低的延迟，并且可以集成到任何现有的搜索端点中，而无需维护向量数据库。让我们深入了解一下！我们将使用 News API 作为示例域进行搜索。

设置

除了您的 OPENAI_API_KEY 之外，您还需要在您的环境中包含 NEWS_API_KEY。您可以在这里获取 API 密钥。

# Dependencies from datetime import date, timedelta # date handling for fetching recent news from IPython import display # for pretty printing import json # for parsing the JSON api responses and model outputs from numpy import dot # for cosine similarity from openai import OpenAI import os # for loading environment variables import requests # for making the API requests from tqdm.notebook import tqdm # for printing progress bars client = OpenAI(api_key=os.environ.get("OPENAI_API_KEY", "<your OpenAI API key if not set as env var>")) # Load environment variables news_api_key = os.getenv("NEWS_API_KEY") GPT_MODEL = "gpt-3.5-turbo" # Helper functions def json_gpt(input: str): completion = client.chat.completions.create(model=GPT_MODEL, messages=[ {"role": "system", "content": "Output only valid JSON"}, {"role": "user", "content": input}, ], temperature=0.5) text = completion.choices[0].message.content parsed = json.loads(text) return parsed def embeddings(input: list[str]) -> list[list[str]]: response = client.embeddings.create(model="text-embedding-3-small", input=input) return [data.embedding for data in response.data]

1. 搜索

一切都始于用户问题。

现在，为了尽可能详尽，我们使用模型根据这个问题生成一个多样化的查询列表。

QUERIES_INPUT = f""" You have access to a search API that returns recent news articles. Generate an array of search queries that are relevant to this question. Use a variation of related keywords for the queries, trying to be as general as possible. Include as many queries as you can think of, including and excluding terms. For example, include queries like ['keyword_1 keyword_2', 'keyword_1', 'keyword_2']. Be creative. The more queries you include, the more likely you are to find relevant results. User question: {USER_QUESTION} Format: {{"queries": ["query_1", "query_2", "query_3"]}} """ queries = json_gpt(QUERIES_INPUT)["queries"] # Let's include the original question as well for good measure queries.append(USER_QUESTION) queries

['NBA championship winner', 'MVP of NBA championship', 'Last game of NBA championship', 'NBA finals winner', 'Most valuable player of NBA championship', 'Finals game of NBA', 'Who won the NBA finals', 'NBA championship game summary', 'NBA finals MVP', 'Champion of NBA playoffs', 'NBA finals last game highlights', 'NBA championship series result', 'NBA finals game score', 'NBA finals game recap', 'NBA champion team and player', 'NBA finals statistics', 'NBA championship final score', 'NBA finals best player', 'NBA playoffs champion and MVP', 'NBA finals game analysis', 'Who won the NBA championship? And who was the MVP? Tell me a bit about the last game.']

查询看起来不错，让我们运行搜索。

def search_news( query: str, news_api_key: str = news_api_key, num_articles: int = 50, from_datetime: str = "2023-06-01", # the 2023 NBA finals were played in June 2023 to_datetime: str = "2023-06-30", ) -> dict: response = requests.get( "https://newsapi.org/v2/everything", params={ "q": query, "apiKey": news_api_key, "pageSize": num_articles, "sortBy": "relevancy", "from": from_datetime, "to": to_datetime, }, ) return response.json() articles = [] for query in tqdm(queries): result = search_news(query) if result["status"] == "ok": articles = articles + result["articles"] else: raise Exception(result["message"]) # remove duplicates articles = list({article["url"]: article for article in articles}.values()) print("Total number of articles:", len(articles)) print("Top 5 articles of query 1:", "\n") for article in articles[0:5]: print("Title:", article["title"]) print("Description:", article["description"]) print("Content:", article["content"][0:100] + "...") print()

Total number of articles: 554 Top 5 articles of query 1: Title: Nascar takes on Le Mans as LeBron James gets centenary race under way Description: <ul><li>Nascar has presence at iconic race for first time since 1976</li><li>NBA superstar LeBron James waves flag as honorary starter</li></ul>The crowd chanted “U-S-A! U-S-A!” as Nascar driver lineup for the 24 Hours of Le Mans passed through the city cente… Content: The crowd chanted U-S-A! U-S-A! as Nascar driver lineup for the 24 Hours of Le Mans passed through t... Title: NBA finals predictions: Nuggets or Heat? Our writers share their picks Description: Denver or Miami? Our contributors pick the winner, key players and dark horses before the NBA’s grand finale tips offA lot has been made of the importance of a balanced roster with continuity, but, somehow, still not enough. The Nuggets are the prime example … Content: The Nuggets are here because A lot has been made of the importance of a balanced roster with conti... Title: Unboxing: Michelob ULTRA and Artist Futura Enshrine the NBA Championship In Custom Hand-Painted Bottles Description: As the 2022-2023 NBA Championship nears the end, Michelob ULTRA brings joy to sports fans who will gather to watch the showdown between the Denver Nuggets and Miami Heat. The beermaker teamed up with artist Futura to remix its newly-designed 2023 Champ Bottle… Content: As the 2022-2023 NBA Championship nears the end, Michelob ULTRA brings joy to sports fans who will g... Title: Futura and Michelob ULTRA Toast to the NBA Finals With Abstract Artwork Crafted From the Brand’s 2023 Limited-Edition Championship Bottles Description: The sun is out to play, and so is Michelob ULTRA. With the 2022-2023 NBA Finals underway, the beermaker is back with its celebratory NBA Champ Bottles. This year, the self-proclaimed MVP of joy is dropping a limited-edition bottle made in collaboration with a… Content: The sun is out to play, and so is Michelob ULTRA. With the 2022-2023 NBA Finals underway, the beerma... Title: Signed and Delivered, Futura and Michelob ULTRA Will Gift Hand-Painted Bottles to This Year’s NBA Championship Team Description: Michelob ULTRA, the MVP of joy and official beer sponsor of the NBA is back to celebrate with basketball lovers and sports fans around the globe as the NBA 2022-2023 season comes to a nail-biting close. In collaboration with artist Futura, Michelob ULTRA will… Content: Michelob ULTRA, the MVP of joy and official beer sponsor of the NBA is back to celebrate with basket...

正如我们所看到的，通常，搜索查询会返回大量结果，其中许多结果与用户提出的原始问题无关。为了提高最终答案的质量，我们使用嵌入来重新排序和筛选结果。

2. 重新排序

从 HyDE (Gao et al.) 中汲取灵感，我们首先生成一个理想的假设答案，以便对我们的结果进行重新排序和比较。这有助于优先考虑看起来像好答案的结果，而不是那些与我们的问题相似的结果。以下是我们用于生成假设答案的提示。

HA_INPUT = f""" Generate a hypothetical answer to the user's question. This answer will be used to rank search results. Pretend you have all the information you need to answer, but don't use any actual facts. Instead, use placeholders like NAME did something, or NAME said something at PLACE. User question: {USER_QUESTION} Format: {{"hypotheticalAnswer": "hypothetical answer text"}} """ hypothetical_answer = json_gpt(HA_INPUT)["hypotheticalAnswer"] hypothetical_answer

'The NBA championship was won by TEAM NAME. The MVP was awarded to PLAYER NAME. The last game was held at STADIUM NAME, where both teams played with great energy and enthusiasm. It was a close game, but in the end, TEAM NAME emerged victorious.'

现在，让我们为搜索结果和假设答案生成嵌入。然后我们计算这些嵌入之间的余弦距离，从而得到语义相似性度量。请注意，由于 OpenAI 嵌入在我们的 API 中返回时已归一化，因此我们可以简单地计算点积来代替执行完整的余弦相似度计算。

hypothetical_answer_embedding = embeddings(hypothetical_answer)[0] article_embeddings = embeddings( [ f"{article['title']} {article['description']} {article['content'][0:100]}" for article in articles ] ) # Calculate cosine similarity cosine_similarities = [] for article_embedding in article_embeddings: cosine_similarities.append(dot(hypothetical_answer_embedding, article_embedding)) cosine_similarities[0:10]

[0.7854456526852069, 0.8086023500072106, 0.8002998147018501, 0.7961229569526956, 0.798354506673743, 0.758216458795653, 0.7753754083127359, 0.7494958338411927, 0.804733946801739, 0.8405965885235218]

最后，我们使用这些相似度分数对结果进行排序和筛选。

scored_articles = zip(articles, cosine_similarities) # Sort articles by cosine similarity sorted_articles = sorted(scored_articles, key=lambda x: x[1], reverse=True) # Print top 5 articles print("Top 5 articles:", "\n") for article, score in sorted_articles[0:5]: print("Title:", article["title"]) print("Description:", article["description"]) print("Content:", article["content"][0:100] + "...") print("Score:", score) print()

Top 5 articles: Title: NBA Finals: Denver Nuggets beat Miami Hea, lift thier first-ever NBA title Description: Denver Nuggets won their maiden NBA Championship trophy defeating Miami Heat 94-89 in Game 5 of the NBA Final held on Tuesday at the Ball Arena in Denver Content: Denver Nuggets won their maiden NBA Championship trophy defeating Miami Heat 94-89 in Game 5 of the ... Score: 0.8445817523602124 Title: Photos: Denver Nuggets celebrate their first NBA title Description: The Nuggets capped off an impressive postseason by beating the Miami Heat in the NBA Finals. Content: Thousands of supporters watched along the streets of Denver, Colorado as the US National Basketball ... Score: 0.842070667753606 Title: Denver Nuggets win first NBA championship title in Game 5 victory over Miami Heat Description: The Denver Nuggets won their first NBA championship Monday night, downing the Miami Heat 94-89 at Ball Arena in Denver to take Game 5 of the NBA Finals. Content: The Denver Nuggets won their first NBA championship Monday night, downing the Miami Heat 94-89 at Ba... Score: 0.8409346078172385 Title: Denver Nuggets Capture Their First NBA Championship Behind Unbreakable Chemistry Description: After 47 years of waiting, the Denver Nuggets are NBA champions. Led by Nikola Jokic and Jamal Murray, they reached the mountain top by staying true to themselves. Content: DENVER, CO - JUNE 12: Jamal Murray (27) of the Denver Nuggets celebrates as he leaves the court ... ... Score: 0.8405965885235218 Title: NBA Finals: Nikola Jokic, Denver Nuggets survive Miami Heat to secure franchise's first NBA championship Description: In a rock-fight of a Game 5, the Denver Nuggets reached the NBA mountaintop from the foothills of the Rockies, winning their first-ever championship and setting Nikola Jokic's legacy as an all-timer in stone. Content: DENVER, COLORADO - JUNE 12: Jamal Murray #27 of the Denver Nuggets reacts during the fourth quarter ... Score: 0.8389716330890262

太棒了！这些结果看起来与我们的原始查询更相关。现在，让我们使用前 5 个结果生成最终答案。

3. 回答

formatted_top_results = [ { "title": article["title"], "description": article["description"], "url": article["url"], } for article, _score in sorted_articles[0:5] ] ANSWER_INPUT = f""" Generate an answer to the user's question based on the given search results. TOP_RESULTS: {formatted_top_results} USER_QUESTION: {USER_QUESTION} Include as much information as possible in the answer. Reference the relevant search result urls as markdown links. """ completion = client.chat.completions.create( model=GPT_MODEL, messages=[{"role": "user", "content": ANSWER_INPUT}], temperature=0.5, stream=True, ) text = "" for chunk in completion: text += chunk.choices[0].delta.content display.clear_output(wait=True) display.display(display.Markdown(text))