Merge pull request #3 from CCCBora/pre-defined-references

app.py

@@ -6,18 +6,21 @@ from utils.file_operations import hash_name
 
 # note: App白屏bug：允许第三方cookie
 # todo:
-#   5. Use some simple method for simple tasks
-#   (including: writing abstract, conclusion, generate keywords, generate figures...)
-#       5.1 Use GPT 3.5 for abstract, conclusion, ... (or may not)
-#       5.2 Use local LLM to generate keywords, figures, ...
-#       5.3 Use embedding to find most related papers (find a paper dataset)
-#   6. get logs when the procedure is not completed.
+#   6. get logs when the procedure is not completed. *
 #   7. 自己的文件库； 更多的prompts
-#   11. distinguish citep and citet
+#   8. Decide on how to generate the main part of a paper * (Langchain/AutoGPT
+#   9. Load .bibtex file to generate a pre-defined references list. *
+#   1. 把paper改成纯JSON?
+#   2. 实现别的功能
+#   3. Check API Key GPT-4 Support.
+#   8. Re-build some components using `langchain`
+#           - in `references.py`, use PromptTemplates.format -> str
+#           - in `gpt_interation`, use LLM
 # future:
-#   8. Change prompts to langchain
 #   4. add auto_polishing function
 #   12. Change link to more appealing color # after the website is built;
+#   1. Check if there are any duplicated citations
+#   2. Remove potential thebibliography and bibitem in .tex file
 
 openai_key = os.getenv("OPENAI_API_KEY")
 access_key_id = os.getenv('AWS_ACCESS_KEY_ID')
@@ -109,19 +112,77 @@ with gr.Blocks(theme=theme) as demo:
     
     输入想要生成的论文名称（比如Playing Atari with Deep Reinforcement Learning), 点击Submit, 等待大概十分钟, 下载.zip格式的输出，在Overleaf上编译浏览.  
     ''')
+
     with gr.Row():
         with gr.Column(scale=2):
             key = gr.Textbox(value=openai_key, lines=1, max_lines=1, label="OpenAI Key",
                              visible=not IS_OPENAI_API_KEY_AVAILABLE)
+
             # generator = gr.Dropdown(choices=["学术论文", "文献总结"], value="文献总结",
             # label="Selection", info="目前支持生成'学术论文'和'文献总结'.", interactive=True)
-            title = gr.Textbox(value="Playing Atari with Deep Reinforcement Learning", lines=1, max_lines=1,
-                               label="Title", info="论文标题")
-            description = gr.Textbox(lines=5, label="Description (Optional)", visible=False)
 
-            with gr.Row():
-                clear_button = gr.Button("Clear")
-                submit_button = gr.Button("Submit", variant="primary")
+            # 每个功能做一个tab
+            with gr.Tab("学术论文"):
+                title = gr.Textbox(value="Playing Atari with Deep Reinforcement Learning", lines=1, max_lines=1,
+                                   label="Title", info="论文标题")
+
+                with gr.Accordion("高级设置", open=False):
+                    description_pp = gr.Textbox(lines=5, label="Description (Optional)", visible=True,
+                                                info="对希望生成的论文的一些描述. 包括这篇论文的创新点, 主要贡献, 等.")
+
+                    interactive = False
+                    gr.Markdown('''
+                    ## 下面的功能我只做了UI, 还没来得及实现功能. 
+                    ''')
+                    with gr.Row():
+                        with gr.Column():
+                            gr.Markdown('''
+                            Upload .bib file (Optional)
+                            
+                            通过上传.bib文件来控制GPT-4模型必须参考哪些文献. 
+                            ''')
+                            bibtex_file = gr.File(label="Upload .bib file", file_types=["text"],
+                                                  interactive=interactive)
+                        with gr.Column():
+                            search_engine = gr.Dropdown(label="Search Engine",
+                                                        choices=["ArXiv", "Semantic Scholar", "Google Scholar", "None"],
+                                                        value= "Semantic Scholar",
+                                                        interactive=interactive,
+                                                        info="用于决定GPT-4用什么搜索引擎来搜索文献. 选择None的时候仅参考给定文献.")
+                            tldr = gr.Checkbox(value=True, label="TLDR;",
+                                               info="选择此筐表示将使用Semantic Scholar的TLDR作为文献的总结.",
+                                               interactive = interactive),
+                            use_cache = gr.Checkbox(label="总是重新生成",
+                                                    info="选择此筐表示将不会读取已经生成好的文章.",
+                                               interactive = interactive)
+                            slider = gr.Slider(minimum=1, maximum=30, value=20, label="最大参考文献数目",
+                                               info="过多参考文献会超出Token数限制导致报错，这里限制最大参考文献数目.")
+
+                with gr.Row():
+                    clear_button_pp = gr.Button("Clear")
+                    submit_button_pp = gr.Button("Submit", variant="primary")
+
+            with gr.Tab("文献综述"):
+                gr.Markdown('''
+                <h1  style="text-align: center;">Coming soon!</h1>
+                ''')
+                # topic = gr.Textbox(value="Deep Reinforcement Learning", lines=1, max_lines=1,
+                #                    label="Topic", info="文献主题")
+                # with gr.Accordion("Advanced Setting"):
+                #     description_lr = gr.Textbox(lines=5, label="Description (Optional)", visible=True,
+                #                              info="对希望生成的综述的一些描述. 包括这篇论文的创新点, 主要贡献, 等.")
+                # with gr.Row():
+                #     clear_button_lr = gr.Button("Clear")
+                #     submit_button_lr = gr.Button("Submit", variant="primary")
+            with gr.Tab("论文润色"):
+                gr.Markdown('''
+                <h1  style="text-align: center;">Coming soon!</h1>
+                ''')
+            with gr.Tab("帮我想想该写什么论文!"):
+                gr.Markdown('''
+                <h1  style="text-align: center;">Coming soon!</h1>
+                ''')
+
         with gr.Column(scale=1):
             style_mapping = {True: "color:white;background-color:green",
                              False: "color:white;background-color:red"}  # todo: to match website's style
@@ -133,8 +194,8 @@ with gr.Blocks(theme=theme) as demo:
             `OpenAI API`: <span style="{style_mapping[IS_OPENAI_API_KEY_AVAILABLE]}">{availability_mapping[IS_OPENAI_API_KEY_AVAILABLE]}</span>.  `Cache`: <span style="{style_mapping[IS_CACHE_AVAILABLE]}">{availability_mapping[IS_CACHE_AVAILABLE]}</span>.''')
             file_output = gr.File(label="Output")
 
-    clear_button.click(fn=clear_inputs, inputs=[title, description], outputs=[title, description])
-    submit_button.click(fn=wrapped_generator, inputs=[title, description, key], outputs=file_output)
+    clear_button_pp.click(fn=clear_inputs, inputs=[title, description_pp], outputs=[title, description_pp])
+    submit_button_pp.click(fn=wrapped_generator, inputs=[title, description_pp, key], outputs=file_output)
 
 demo.queue(concurrency_count=1, max_size=5, api_open=False)
 demo.launch()

auto_backgrounds.py

@@ -91,7 +91,7 @@ def fake_generator(title, description="", template="ICLR2022", model="gpt-4"):
     return make_archive("sample-output.pdf", filename)
 
 
-def generate_draft(title, description="", template="ICLR2022", model="gpt-4", search_engine="ss", tldr=True, max_kw_refs=14):
+def generate_draft(title, description="", template="ICLR2022", model="gpt-4", search_engine="ss", tldr=True, max_kw_refs=10):
     paper, destination_folder, _ = _generation_setup(title, description, template, model, search_engine, tldr, max_kw_refs)
 
     # todo: `list_of_methods` failed to be generated; find a solution ...

latex_templates/pre_refs.bib

@@ -0,0 +1,17 @@
+
+@article{1512.07669,
+          title = {Reinforcement Learning: Stochastic Approximation Algorithms for Markov
+  Decision Processes},
+          author = {Vikram Krishnamurthy},
+          journal={arXiv preprint arXiv:1512.07669},
+          year = {2015},
+          url = {http://arxiv.org/abs/1512.07669v1}
+        }
+
+@article{1511.02377,
+          title = {The Value Functions of Markov Decision Processes},
+          author = {Ehud Lehrer , Eilon Solan , Omri N. Solan},
+          journal={arXiv preprint arXiv:1511.02377},
+          year = {2015},
+          url = {http://arxiv.org/abs/1511.02377v1}
+        }

utils/references.py

@@ -1,18 +1,26 @@
-# Generate references
-#   1. select most correlated references from "references" dataset or Arxiv search engine.
-#   2. Generate bibtex from the selected papers. --> to_bibtex()
-#   3. Generate prompts from the selected papers: --> to_prompts()
-#       {"paper_id": "paper summary"}
-
+# Each `paper` is a dictionary containing:
+#       (1) paper_id (2) title (3) authors (4) year (5) link (6) abstract (7) journal
+#
+# Generate references:
+#   `Reference` class:
+#       1. Read a given .bib file to collect papers; use `search_paper_abstract` method to fill missing abstract.
+#       2. Given some keywords; use ArXiv or Semantic Scholar API to find papers.
+#       3. Generate bibtex from the selected papers. --> to_bibtex()
+#       4. Generate prompts from the selected papers: --> to_prompts()
+#               A sample prompt: {"paper_id": "paper summary"}
 
 import requests
 import re
+import bibtexparser
+from scholarly import scholarly
+from scholarly import ProxyGenerator
 
 
 ######################################################################################################################
 # Some basic tools
 ######################################################################################################################
 def remove_newlines(serie):
+    # This function is applied to the abstract of each paper to reduce the length of prompts.
     serie = serie.replace('\n', ' ')
     serie = serie.replace('\\n', ' ')
     serie = serie.replace('  ', ' ')
@@ -20,6 +28,47 @@ def remove_newlines(serie):
     return serie
 
 
+def search_paper_abstract(title):
+    pg = ProxyGenerator()
+    success = pg.ScraperAPI("921b16f94d701308b9d9b4456ddde155")
+    scholarly.use_proxy(pg)
+    # input the title of a paper, return its abstract
+    search_query = scholarly.search_pubs(title)
+    paper = next(search_query)
+    return remove_newlines(paper['bib']['abstract'])
+
+
+def load_papers_from_bibtex(bib_file_path):
+    with open(bib_file_path) as bibtex_file:
+        bib_database = bibtexparser.load(bibtex_file)
+    if len(bib_database.entries) == 0:
+        return []
+    else:
+        bib_papers = []
+        for bibitem in bib_database.entries:
+            paper_id = bibitem.get("ID")
+            title = bibitem.get("title")
+            if title is None:
+                continue
+            journal = bibitem.get("journal")
+            year = bibitem.get("year")
+            author = bibitem.get("author")
+            abstract = bibitem.get("abstract")
+            if abstract is None:
+                abstract = search_paper_abstract(title)
+            result = {
+                "paper_id": paper_id,
+                "title": title,
+                "link": "",
+                "abstract": abstract,
+                "authors": author,
+                "year": year,
+                "journal": journal
+            }
+            bib_papers.append(result)
+        return bib_papers
+
+
 ######################################################################################################################
 # Semantic Scholar (SS) API
 ######################################################################################################################
@@ -63,7 +112,11 @@ def _collect_papers_ss(keyword, counts=3, tldr=False):
         pattern = r'^\w+'
         words = re.findall(pattern, title)
         # return last_name + year_str + title.split(' ', 1)[0]
-        return last_name + year_str + words[0]
+        try:
+            output = last_name + year_str + words[0]
+        except IndexError:
+            output = last_name + year_str + title[:4]
+        return output
 
     def extract_author_info(raw_authors):
         authors = [author['name'] for author in raw_authors]
@@ -71,7 +124,7 @@ def _collect_papers_ss(keyword, counts=3, tldr=False):
         authors_str = " and ".join(authors)
         try:
             last_name = authors[0].split()[-1]
-        except:
+        except IndexError:
             last_name = "ma"
         # pattern = r'^\w+'
         # last_name = re.findall(pattern, authors[0])
@@ -79,7 +132,7 @@ def _collect_papers_ss(keyword, counts=3, tldr=False):
 
     def parse_search_results(search_results_ss):
         # turn the search result to a list of paper dictionary.
-        papers = []
+        papers_ss = []
         for raw_paper in search_results_ss:
             if raw_paper["abstract"] is None:
                 continue
@@ -100,14 +153,14 @@ def _collect_papers_ss(keyword, counts=3, tldr=False):
             result = {
                 "paper_id": paper_id,
                 "title": title,
-                "abstract": abstract,  # todo: compare results with tldr
+                "abstract": abstract,
                 "link": link,
                 "authors": authors_str,
                 "year": year_str,
                 "journal": journal
             }
-            papers.append(result)
-        return papers
+            papers_ss.append(result)
+        return papers_ss
 
     raw_results = ss_search(keyword, limit=counts)
     if raw_results is not None:
@@ -192,13 +245,13 @@ def _collect_papers_arxiv(keyword, counts=3, tldr=False):
 # References Class
 ######################################################################################################################
 
-# Each `paper` is a dictionary containing (1) paper_id (2) title (3) authors (4) year (5) link (6) abstract (7) journal
 class References:
     def __init__(self, load_papers=""):
         if load_papers:
-            # todo: read a json file from the given path
-            #       this could be used to support pre-defined references
-            pass
+            # todo: (1) too large bibtex may make have issues on token limitations; may truncate to 5 or 10
+            #       (2) google scholar didn't give a full abstract for some papers ...
+            #       (3) may use langchain to support long input
+            self.papers = load_papers_from_bibtex(load_papers)
         else:
             self.papers = []
 
@@ -266,15 +319,20 @@ class References:
 
 
 if __name__ == "__main__":
-    refs = References()
-    keywords_dict = {
-        "Deep Q-Networks": 15,
-        "Policy Gradient Methods": 24,
-        "Actor-Critic Algorithms": 4,
-        "Model-Based Reinforcement Learning": 13,
-        "Exploration-Exploitation Trade-off": 7
-    }
-    refs.collect_papers(keywords_dict, method="ss", tldr=True)
-    for p in refs.papers:
-        print(p["paper_id"])
-    print(len(refs.papers))
+    # refs = References()
+    # keywords_dict = {
+    #     "Deep Q-Networks": 15,
+    #     "Policy Gradient Methods": 24,
+    #     "Actor-Critic Algorithms": 4,
+    #     "Model-Based Reinforcement Learning": 13,
+    #     "Exploration-Exploitation Trade-off": 7
+    # }
+    # refs.collect_papers(keywords_dict, method="ss", tldr=True)
+    # for p in refs.papers:
+    #     print(p["paper_id"])
+    # print(len(refs.papers))
+
+    bib = "D:\\Projects\\auto-draft\\latex_templates\\pre_refs.bib"
+    papers = load_papers_from_bibtex(bib)
+    for paper in papers:
+        print(paper)

utils/tex_processing.py

@@ -2,16 +2,12 @@ import os
 
 def replace_title(save_to_path, title):
     # Define input and output file names
-    # input_file_name =  save_to_path + "/template.tex"
-    # output_file_name = save_to_path + "/main.tex"
     input_file_name = os.path.join(save_to_path, "template.tex")
     output_file_name = os.path.join(save_to_path , "main.tex")
 
     # Open the input file and read its content
     with open(input_file_name, 'r') as infile:
         content = infile.read()
-
-    # Replace all occurrences of "asdfgh" with "hahaha"
     content = content.replace(r"\title{TITLE} ", f"\\title{{{title}}} ")
 
     # Open the output file and write the modified content
@@ -19,11 +15,14 @@ def replace_title(save_to_path, title):
         outfile.write(content)
 
 
-# return all string in \cite{...}.
+# return all string in \cite{...} \citet{...} or \citep{...}.
 
 # check if citations are in bibtex.
 
 
 # replace citations
 
-# sometimes the output may include thebibliography and bibitem . remove all of it.
+# sometimes the output may include thebibliography and bibitem . remove all of it.
+
+
+