From 83f0996ec9913d7a5a9d5829caad9cb723f6e72b Mon Sep 17 00:00:00 2001 From: Pratyay Pandey Date: Fri, 28 Feb 2025 15:07:37 -0800 Subject: [PATCH] Add deep research demo temporarily as a py file --- .../deep_research_demo/deep_research_demo.py | 101 ++++++++++++++++++ 1 file changed, 101 insertions(+) create mode 100644 examples/deep_research_demo/deep_research_demo.py diff --git a/examples/deep_research_demo/deep_research_demo.py b/examples/deep_research_demo/deep_research_demo.py new file mode 100644 index 00000000..6b44ab06 --- /dev/null +++ b/examples/deep_research_demo/deep_research_demo.py @@ -0,0 +1,101 @@ +import os + +import pandas as pd + +import lotus +from lotus import WebSearchCorpus, parse_pdf, web_search +from lotus.models import LM + +os.environ["SERPAPI_API_KEY"] = "" +os.environ["OPENAI_API_KEY"] = "" + +lm = LM(model="gpt-4o-mini") + +lotus.settings.configure(lm=lm) + +""" +Creating a deep research pipeline with LOTUS is as simple as two commands -- `web_search` and `sem_agg`. +We can use `web_search` to pull a list of relevant articles from Google. From here, we can use `sem_agg` to +condense the articles into a more concise response. This directly makes use of Internet sources to make the +most accurate and up-to-date response. +""" + +question = "Why do attention-based models like Transformers scale better than CNNs or RNNs for large datasets?" +# Pull the first five relevant articles from Google +df_google = web_search(WebSearchCorpus.GOOGLE, question, 5)[["title", "snippet"]] +print(f"Results from Google\n{df_google}") + +# Condense or summarize the articles +df_agg_result = df_google.sem_agg( + "Summarize each {snippet} into a brief paragraph, answering the question: " + question +) +print(f"Summarized results\n{df_agg_result._output[0]}") + +""" +We can also use `web_search` with arXiv to pull a list of relevant research papers for more accurate research. +Let's try to answer the same question once again, but using both Google and arXiv. +""" + +df_arxiv = web_search(WebSearchCorpus.ARXIV, question, 5)[ + ["title", "abstract"] +] # Working with arXiv, we parse the `abstract` of a paper instead of `snippet`. +print(f"Results from arXiv\n{df_arxiv}") + +df_arxiv = df_arxiv.rename(columns={"abstract": "snippet"}) +df_altogether = pd.concat([df_google, df_arxiv]) +df_joint_agg_result = df_altogether.sem_agg( + "Summarize each {snippet} into a brief paragraph, answering the question: " + question +) +print(f"Summarized results\n{df_joint_agg_result._output[0]}") + +""" +Sometimes, less is more. Not all these sources will be equally relevant to the question. Now, let's expand to more than just `web_search` and `sem_agg` +to see if we can filter by relevance. Let's only take the top 5 results out of a combined +DataFrame of fifty Google and arXiv results. This will give us a good blend of +diverse, specific, and relevant sources. +""" + +df_google_twentyfive = web_search(WebSearchCorpus.GOOGLE, question, 25)[["title", "snippet"]] +df_arxiv_twentyfive = web_search(WebSearchCorpus.ARXIV, question, 25)[["title", "abstract"]] +df_arxiv_twentyfive = df_arxiv_twentyfive.rename(columns={"abstract": "snippet"}) + +df_twentyfive = pd.concat([df_google_twentyfive, df_arxiv_twentyfive]) +""" +Use `sem_topk` to filter by relevance. We'll use the `quick` method, which allows us to quickly sort by relevance. +We set `return_stats` to False because we do not require the specific statistics that were measured for scoring. +""" +df_twentyfive = df_twentyfive.sem_topk( + "Which {snippet} is most relevant to the question and unique?", K=5, method="quick", return_stats=False +) + +summarized_answer = df_twentyfive.sem_agg( + "Summarize each {snippet} into a brief paragraph, answering the question: " + question +)._output[0] +print(f"Summarized results\n{summarized_answer}") + +""" +A key advantage to using LOTUS, compared to LLM chats, is also that you possess all the data that you've used to answer the question. +We can continue using `df_twentyfive` to answer follow-up questionss, extracting different pieces of information or builiding on our +repository of context. Now, let's make use of the `parse_pdf` method to also pass in a textbook relevant to this question. +""" + + +""" +Even with unusual formatting or interesting PDF layouts and over multiple pages, `parse_pdf`can accurately parse text rapdily. +We can also simply pass in the URL of the PDF to parse it, so you do not even need to download the file. +""" +attention_is_all_you_need = parse_pdf("https://arxiv.org/pdf/1706.03762", per_page=False).rename( + columns={"file_path": "title", "content": "snippet"} +) +the_transformer = parse_pdf("https://web.stanford.edu/~jurafsky/slp3/9.pdf", per_page=False).rename( + columns={"file_path": "title", "content": "snippet"} +) + +print(f"Parsed 'Attention is All You Need'\n{attention_is_all_you_need}") +print(f"Parsed Jurafsky's 'Speech and Language Processing'\n{the_transformer}") +df_combined = pd.concat([df_twentyfive, attention_is_all_you_need, the_transformer]) + +answer_with_textbooks = df_combined.sem_agg( + "Summarize each {snippet} into a brief paragraph, answering the question: " + question +)._output[0] +print(f"Answer with textbooks\n{answer_with_textbooks}")