Update database

arxivdb/chroma.sqlite3

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0c48f817474996b45a3f4da1e127a2fde083db4bfeddb71893d598b8200fb056
+oid sha256:12ccbdc4e184a9a32b70f35d7f9b7e67a5371b0f7f5514258652eb9e8d71a055
 size 123736064

mainflow_single_question.ipynb

@@ -2,9 +2,18 @@
  "cells": [
   {
    "cell_type": "code",
-   "execution_count": 4,
+   "execution_count": 1,
    "metadata": {},
-   "outputs": [],
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "C:\\Users\\Admin\\AppData\\Local\\Packages\\PythonSoftwareFoundation.Python.3.10_qbz5n2kfra8p0\\LocalCache\\local-packages\\Python310\\site-packages\\tqdm\\auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
+      "  from .autonotebook import tqdm as notebook_tqdm\n"
+     ]
+    }
+   ],
    "source": [
     "import google.generativeai as genai\n",
     "import utils\n",
@@ -15,7 +24,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": null,
+   "execution_count": 6,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -30,16 +39,18 @@
     "            \"Gemini API Key not provided. Please provide GEMINI_API_KEY as an environment variable\"\n",
     "        )\n",
     "    genai.configure(api_key=gemini_api_key)\n",
-    "    model = genai.GenerativeModel(\"gemini-pro\")\n",
+    "    config = genai.GenerationConfig(max_output_tokens=1024,\n",
+    "                                    temperature=0.5)\n",
+    "    model = genai.GenerativeModel(\"gemini-pro\",\n",
+    "                                  generation_config=config)\n",
     "    return model, sqldb, db\n",
     "\n",
-    "\n",
     "model, sqldb, db = configure()"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 2,
+   "execution_count": 12,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -47,7 +58,7 @@
     "    \"\"\"A prompt that return a JSON block as arguments for querying database\"\"\"\n",
     "\n",
     "    prompt = (\n",
-    "        \"\"\"[INST] You are an assistant that choose only one action below based on guest question.\n",
+    "        \"\"\"[INST] SYSTEM: You are an assistant that choose only one action below based on guest question.\n",
     "   1. If the guest question is asking for some specific document or article, you need to respond the information in JSON format with 2 keys \"title\", \"author\" if found any above. The authors are separated with the word 'and'. \n",
     "   2. If the guest question is asking for relevant informations about a topic, you need to respond the information in JSON format with 2 keys \"keywords\", \"description\", include a list of keywords represent the main academic topic, \\\n",
     "     and a description about the main topic. You may paraphrase the keywords to add more. \\\n",
@@ -66,7 +77,7 @@
     "    prompt = (\n",
     "        \"\"\"[INST] You are an library assistant that help to search articles and documents based on user's question.\n",
     "  From guest's question, you have found some records and documents that may help. Now you need to answer the guest with the information found.\n",
-    "  You should answer in a conversational form politely.\n",
+    "  If no information found in the database, you may generate some other recommendation related to user's question using your own knowledge. You should answer in a conversational form politely.\n",
     "  QUESTION: '{input}'\n",
     "  INFORMATION: '{contexts}'\n",
     "  [/INST]\n",
@@ -79,7 +90,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 3,
+   "execution_count": 8,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -96,6 +107,15 @@
     "        results = db.query_relevant(keywords=keywords, query_texts=query_texts)\n",
     "        # print(results)\n",
     "        ids = results['metadatas'][0]\n",
+    "        if len(ids) == 0:\n",
+    "            # go crawl some\n",
+    "            new_records = utils.crawl_arxiv(keyword_list=keywords)\n",
+    "            if type(new_records) == str:\n",
+    "                return \"Error occured, information not found\", new_records\n",
+    "            db.add(new_records)\n",
+    "            sqldb.add(new_records)\n",
+    "            results = db.query_relevant(keywords=keywords, query_texts=query_texts)\n",
+    "            ids = results['metadatas'][0]\n",
     "        paper_id = [id['paper_id'] for id in ids]\n",
     "        paper_info = sqldb.query_id(paper_id)\n",
     "        # print(paper_info)\n",
@@ -109,16 +129,20 @@
     "        # invoke llm and return result\n",
     "\n",
     "    if \"title\" in keys:\n",
-    "        results = sqldb.query(title = args['title'],author = args['author'])\n",
-    "        print(results)\n",
-    "        paper_info = sqldb.query(title = args['title'],author = args['author'])\n",
+    "        title = args['title']\n",
+    "        authors = utils.process_authors_list(args['author'])\n",
+    "        paper_info = sqldb.query(title = title,author = authors)\n",
     "        # if query not found then go crawl brh\n",
+    "        # print(paper_info)\n",
     "\n",
     "        if len(paper_info) == 0:\n",
-    "            new_records = utils.crawl_exact_paper(title=args['title'],author=args['author'])\n",
+    "            new_records = utils.crawl_exact_paper(title=title,author=authors)\n",
+    "            if type(new_records) == str:\n",
+    "                # print(new_records)\n",
+    "                return \"Error occured, information not found\", \"Information not found\"\n",
     "            db.add(new_records)\n",
     "            sqldb.add(new_records)\n",
-    "            paper_info = sqldb.query(title = args['title'],author = args['author'])\n",
+    "            paper_info = sqldb.query(title = title,author = authors)\n",
     "        # -------------------------------------\n",
     "        records = [] # get title (2), author (3), link (6)\n",
     "        result_string = \"\"\n",
@@ -127,14 +151,14 @@
     "            records.append([paper_info[i][2],paper_info[i][3],paper_info[i][6]])\n",
     "        # process results:\n",
     "        if len(result_string) == 0:\n",
-    "            return \"Information not found\", None\n",
+    "            return \"Information not found\", \"Information not found\"\n",
     "        return result_string, records\n",
     "        # invoke llm and return result"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 8,
+   "execution_count": 5,
    "metadata": {},
    "outputs": [
     {
@@ -143,27 +167,27 @@
      "text": [
       "--------------------------\n",
       "{\n",
-      " \"keywords\": [\"Video action recognition\", \"Long Short-Term Memory\", \"Computer vision\", \"Deep learning\", \"Time series analysis\"],\n",
-      " \"description\": \"Video action recognition is a challenging task in computer vision. It requires the ability to understand the temporal dynamics of a video and to identify the actions being performed. Long Short-Term Memory (LSTM) networks are a type of recurrent neural network that is well-suited for learning long-term dependencies. This makes them a good choice for video action recognition tasks. LSTM networks have been shown to achieve state-of-the-art results on a variety of video action recognition datasets using deep learning techniques.\"\n",
+      " \"keywords\": [\"Long short-term memory\", \"Video action recognition\", \"Convolutional neural networks\"],\n",
+      " \"description\": \"Long short-term memory (LSTM) is a type of recurrent neural network (RNN) that is used for processing sequential data. It is particularly well-suited for tasks such as video action recognition, where the order of the frames in the video is important. Convolutional neural networks (CNNs) are another type of neural network that is often used for image and video processing. They are particularly good at learning the spatial features of images and videos. LSTM networks and CNNs can be used together to create powerful models for video action recognition.\"\n",
       "}\n"
      ]
     },
     {
      "data": {
       "text/plain": [
-       "('Title: Applications of Deep Neural Networks with Keras, Author: Jeff Heato, Link: http://arxiv.org/pdf/2009.05673v5Title: DeepDrum: An Adaptive Conditional Neural Network, Author: Dimos Makris, Maximos Kaliakatsos-Papakostas, Katia Lida Kermanidi, Link: http://arxiv.org/pdf/1809.06127v2Title: A Video Recognition Method by using Adaptive Structural Learning of Long Short Term Memory based Deep Belief Network, Author: Shin Kamada, Takumi Ichimur, Link: http://arxiv.org/pdf/1909.13480v1',\n",
-       " [['Applications of Deep Neural Networks with Keras',\n",
-       "   'Jeff Heato',\n",
-       "   'http://arxiv.org/pdf/2009.05673v5'],\n",
-       "  ['DeepDrum: An Adaptive Conditional Neural Network',\n",
-       "   'Dimos Makris, Maximos Kaliakatsos-Papakostas, Katia Lida Kermanidi',\n",
-       "   'http://arxiv.org/pdf/1809.06127v2'],\n",
-       "  ['A Video Recognition Method by using Adaptive Structural Learning of Long Short Term Memory based Deep Belief Network',\n",
-       "   'Shin Kamada, Takumi Ichimur',\n",
-       "   'http://arxiv.org/pdf/1909.13480v1']])"
+       "('Title: LiteLSTM Architecture for Deep Recurrent Neural Networks, Author: Nelly Elsayed, Zag ElSayed, Anthony S. Maid, Link: http://arxiv.org/pdf/2201.11624v2Title: AdderNet and its Minimalist Hardware Design for Energy-Efficient Artificial Intelligence, Author: Yunhe Wang, Mingqiang Huang, Kai Han, Hanting Chen, Wei Zhang, Chunjing Xu, Dacheng Ta, Link: http://arxiv.org/pdf/2101.10015v2Title: Audio Tagging on an Embedded Hardware Platform, Author: Gabriel Bibbo, Arshdeep Singh, Mark D. Plumble, Link: http://arxiv.org/pdf/2306.09106v1',\n",
+       " [['LiteLSTM Architecture for Deep Recurrent Neural Networks',\n",
+       "   'Nelly Elsayed, Zag ElSayed, Anthony S. Maid',\n",
+       "   'http://arxiv.org/pdf/2201.11624v2'],\n",
+       "  ['AdderNet and its Minimalist Hardware Design for Energy-Efficient Artificial Intelligence',\n",
+       "   'Yunhe Wang, Mingqiang Huang, Kai Han, Hanting Chen, Wei Zhang, Chunjing Xu, Dacheng Ta',\n",
+       "   'http://arxiv.org/pdf/2101.10015v2'],\n",
+       "  ['Audio Tagging on an Embedded Hardware Platform',\n",
+       "   'Gabriel Bibbo, Arshdeep Singh, Mark D. Plumble',\n",
+       "   'http://arxiv.org/pdf/2306.09106v1']])"
       ]
      },
-     "execution_count": 8,
+     "execution_count": 5,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -187,68 +211,51 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 10,
+   "execution_count": 17,
    "metadata": {},
    "outputs": [
     {
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "For recognizing actions in videos using Long Short-Term Memory (LSTM) models, you may find the following papers insightful. These studies explore various aspects of action recognition using LSTM: (1) 'Action Recognition with 3D Convolutional Neural Networks and Long Short-Term Memory' by Tran et al., 2015; (2) 'ConvLSTM: A Convolutional Neural Network for Video Recognition' by Shi et al., 2015; (3) 'Using Long-Short Term Memory for Action Recognition in Videos' by Graves, 2013.\n",
-      "['LSTM model', 'video analysis', 'action recognition']\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "Llama.generate: prefix-match hit\n",
-      "\n",
-      "llama_print_timings:        load time =  139768.70 ms\n",
-      "llama_print_timings:      sample time =     140.16 ms /   412 runs   (    0.34 ms per token,  2939.41 tokens per second)\n",
-      "llama_print_timings: prompt eval time =       0.00 ms /     1 tokens (    0.00 ms per token,      inf tokens per second)\n",
-      "llama_print_timings:        eval time =   91570.34 ms /   412 runs   (  222.26 ms per token,     4.50 tokens per second)\n",
-      "llama_print_timings:       total time =   93048.98 ms /   413 tokens\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "------------------------\n",
-      "Hello there! I see that you're working on an LSTM model for recognizing actions in videos. I have found some related papers that might be of interest to you.\n",
-      "\n",
-      "  1. The first one is titled \"Action in Mind: A Neural Network Approach to Action Recognition and Segmentation\" by Zahra Gharae. This paper proposes a neural network approach for action recognition and segmentation, which could provide some insights into your work with LSTM models. You can find it at this link: <http://arxiv.org/pdf/2104.14870v1>\n",
-      "\n",
-      "  2. Another interesting paper is \"Deep Neural Networks in Video Human Action Recognition: A Review\" by Zihan Wang, Yang Yang, Zhi Liu, and Yifan Zhen. This review discusses the application of deep neural networks for video human action recognition. It could give you a broader perspective on the current state-of-the-art methods in this field. You can access it here: <http://arxiv.org/pdf/2305.15692v1>\n",
-      "\n",
-      "  3. Lastly, there's \"3D Convolutional Neural Networks for Ultrasound-Based Silent Speech Interfaces\" by László Tóth and Amin Honarmandi Shandi. Although the title might not seem directly related to your work on LSTM models for action recognition, it does involve the use of 3D convolutional neural networks in video processing, which could still provide valuable insights. You can read it at this link: <http://arxiv.org/pdf/2104.11532v1>\n",
-      "\n",
-      "  I hope you find these resources helpful! Let me know if there's anything else I can assist you with. Have a great day!\n"
+      "near \")\": syntax error\n",
+      "Error query:  select * from arxivsql where id in )\n",
+      "\n"
      ]
     }
    ],
    "source": [
     "# test response, second step\n",
-    "input_prompt = \"I'm working on a LSTM model to recognize actions in a video, recommend me some related papers\"\n",
-    "args =  {\n",
-    "      \"keywords\": [\"LSTM model\", \"video analysis\", \"action recognition\"],\n",
-    "      \"description\": \"For recognizing actions in videos using Long Short-Term Memory (LSTM) models, you may find the following papers insightful. These studies explore various aspects of action recognition using LSTM: (1) 'Action Recognition with 3D Convolutional Neural Networks and Long Short-Term Memory' by Tran et al., 2015; (2) 'ConvLSTM: A Convolutional Neural Network for Video Recognition' by Shi et al., 2015; (3) 'Using Long-Short Term Memory for Action Recognition in Videos' by Graves, 2013.\"\n",
-    "    }\n",
+    "input_prompt = \"Can you suggest some key papers on model predictive control for nonlinear systems, and are there any recent reviews on the application of control theory to robotics?\"\n",
+    "args =  \"{\\n \\\"keywords\\\": [\\n  \\\"Power Electronics\\\",\\n  \\\"Renewable Energy Systems\\\",\\n  \\\"High-Frequency Power Converters\\\",\\n  \\\"Power Electronics Applications\\\",\\n  \\\"Renewable Energy Sources\\\"\\n ],\\n \\\"description\\\": \\\"Power electronics is a branch of electrical engineering that deals with the conversion, control, and conditioning of electrical power. It is used in a wide variety of applications, including power generation, transmission, distribution, and utilization. Renewable energy systems are systems that generate electricity from renewable sources, such as solar, wind, and hydro power. Power electronics is used in renewable energy systems to convert the electrical output of the renewable source into a form that can be used by the grid or by consumers.\\\"\\n}\"\n",
+    "args = json.loads(args)\n",
     "contexts, results = response(args)\n",
     "if not results:\n",
     "  # direct answer\n",
-    "    print(contexts)\n",
+    "  print(contexts)\n",
     "else:\n",
     "  output_prompt = make_answer_prompt(input_prompt,contexts)\n",
-    "  # answer = model.invoke(output_prompt,\n",
-    "  #                       temperature=0.3) # ctrans\n",
-    "  answer = model(prompt=output_prompt,\n",
-    "               temperature=0.0,\n",
-    "               max_tokens=1024,\n",
-    "               ) # llama\n",
-    "  print(\"------------------------\")\n",
-    "  print(answer['choices'][0]['text'])"
+    "  answer = model.generate_content(output_prompt).text\n",
+    "  print(answer)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 14,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "{'desired': 'Natural Language Processing (Computer Science)', 'question': 'What are some recent papers on deep learning architectures for text classification, and can you recommend any surveys or reviews on the topic?'}\n"
+     ]
+    }
+   ],
+   "source": [
+    "with open(\"test_questions.txt\",\"r\") as infile:\n",
+    "    data = json.load(infile)\n",
+    "print(data[0])"
    ]
   },
   {
@@ -287,8 +294,6 @@
    ],
    "source": [
     "# full chain\n",
-    "# inference time depends on hardware ability :')\n",
-    "# with CPU i7-8750H, 16GB RAM and no GPU cuda, expect inference time up to 5-6 min\n",
     "# input_prompt = input()\n",
     "input_prompt = \"I'm working on a LSTM model to recognize actions in a video, recommend me some related papers\"\n",
     "first_prompt = extract_keyword_prompt(input_prompt)\n",
@@ -315,7 +320,61 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 17,
+   "execution_count": 15,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def full_chain_single_question(input_prompt):\n",
+    "    try:\n",
+    "        first_prompt = extract_keyword_prompt(input_prompt)\n",
+    "        temp_answer = model.generate_content(first_prompt).text\n",
+    "\n",
+    "        args = json.loads(utils.trimming(temp_answer))\n",
+    "        contexts, results = response(args)\n",
+    "        if not results:\n",
+    "            print(contexts)\n",
+    "        else:\n",
+    "            output_prompt = make_answer_prompt(input_prompt,contexts)\n",
+    "            answer = model.generate_content(output_prompt).text\n",
+    "        return temp_answer, answer\n",
+    "    except Exception as e:\n",
+    "        return temp_answer, \"Error occured: \" + str(e)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 16,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "[]\n",
+      "[]\n",
+      "Error: timed out\n",
+      "[]\n",
+      "near \")\": syntax error\n",
+      "Error query:  select * from arxivsql where id in )\n",
+      "\n"
+     ]
+    }
+   ],
+   "source": [
+    "test_log = []\n",
+    "for t in data:\n",
+    "    temp_answer, answer = full_chain_single_question(t['question'])\n",
+    "    test_log.append({'desired topic':t['desired'],\n",
+    "           'question':t['question'],\n",
+    "           'first answer':temp_answer,\n",
+    "           'final answer':answer})\n",
+    "with open(\"test_results.json\",\"w\") as outfile:\n",
+    "    json.dump(test_log,outfile)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
    "metadata": {},
    "outputs": [],
    "source": [

utils.py

@@ -27,7 +27,7 @@ class ArxivSQL:
         else:
             query_title = "True"
         if len(author)>0:
-            query_author = 'author like '
+            query_author = 'authors like '
             for auth in author:
                 query_author += "'%{}%' or ".format(auth)
             query_author = query_author[:-4]
@@ -38,12 +38,16 @@ class ArxivSQL:
         return result.fetchall()
 
     def query_id(self, ids=[]):
-        query = "select * from {} where id in (".format(self.table)
-        for id in ids:
-            query+="'"+id+"',"
-        query = query[:-1] + ")"
-        result = self.cur.execute(query)
-        return result.fetchall()
+        try:
+            query = "select * from {} where id in (".format(self.table)
+            for id in ids:
+                query+="'"+id+"',"
+            query = query[:-1] + ")"
+            result = self.cur.execute(query)
+            return result.fetchall()
+        except Exception as e:
+            print(e)
+            print("Error query: ",query)
     
     def add(self, crawl_records):
         """
@@ -71,6 +75,7 @@ class ArxivSQL:
                 result+="\n" + query + "\n"
             finally:
                 return results
+            
 class ArxivChroma:
     """
     Create an interface to arxivdb, which only support query and addition.
@@ -161,7 +166,9 @@ def chunk_text_with_overlap(text, max_char=400, overlap=100):
 def trimming(txt):
     start = txt.find("{")
     end = txt.rfind("}")
-    return txt[start:end+1]
+    return txt[start:end+1].replace("\n"," ")
+
+# crawl data
 
 def extract_tag(txt,tagname):
     return txt[txt.find("<"+tagname+">")+len(tagname)+2:txt.find("</"+tagname+">")]
@@ -194,7 +201,7 @@ def choose_topic(summary):
                                                 cache_dir='models')
     embed = model_embedding.encode(summary)
     cos_sim = lambda a,b: (a @ b.T) / (norm(a)*norm(b))
-    descriptions = json.load(open("topic_descriptions"))
+    descriptions = json.load(open("topic_descriptions.txt"))
     topic = ""
     max_sim = 0.
     for key in descriptions:
@@ -203,21 +210,59 @@ def choose_topic(summary):
             topic = key
             max_sim = sim
     return topic
-        
+
+class TopicClassifier:
+    def __init__(self):
+        self.model_embedding = AutoModel.from_pretrained('jinaai/jina-embeddings-v2-base-en',
+                                                trust_remote_code=True,
+                                                cache_dir='models')
+        topic_descriptions = json.load(open("topic_descriptions.txt")) 
+        self.topics = list(dict.keys(topic_descriptions))
+        self.embeddings = [self.model_embedding.encode(topic_descriptions[key]) for key in topic_descriptions]
+        self.cos_sim = lambda a,b: (a @ b.T) / (norm(a)*norm(b))
+
+    def classifier(self,description):
+        embed = self.model_embedding.encode(description)
+        max_sim = 0.
+        topic = ""
+        for i, key in enumerate(self.topics):
+            sim = self.cos_sim(embed,self.embeddings[i])
+            if sim > max_sim:
+                topic = key
+                max_sim = sim
+        return topic
+
+def crawl_exact_paper(title,author,max_results=3):
+    authors = process_authors_str(author)
+    records = []
+    url = 'http://export.arxiv.org/api/query?search_query=ti:{title}+AND+au:{author}&max_results={max_results}'.format(title=title,author=authors,max_results=max_results)
+    url = url.replace(" ","%20")
+    try:
+        arxiv_page = urllib.request.urlopen(url,timeout=100).read()
+        xml = str(arxiv_page,encoding="utf-8") 
+        while xml.find("<entry>") != -1:
+            extract = xml[xml.find("<entry>")+7:xml.find("</entry>")]
+            xml = xml[xml.find("</entry>")+8:]
+            extract = get_record(extract)
+            topic = choose_topic(extract[6])
+            records.append([topic,*extract])
+        return records
+    except Exception as e:
+        return "Error: "+str(e)
+
 def crawl_arxiv(keyword_list, max_results=100):
     baseurl = 'http://export.arxiv.org/api/query?search_query='
     records = []
-    for keyword in keyword_list:
+    for i,keyword in enumerate(keyword_list):
         if i ==0:
             url = baseurl + 'all:' + keyword
-            i = i + 1               
         else:
             url = url + '+OR+' + 'all:' + keyword
     url = url+ '&max_results=' + str(max_results)
     url = url.replace(' ', '%20')
     try:
         arxiv_page = urllib.request.urlopen(url,timeout=100).read()
-        arxiv_page = str(arxiv_page,encoding="utf-8") 
+        xml = str(arxiv_page,encoding="utf-8") 
         while xml.find("<entry>") != -1:
             extract = xml[xml.find("<entry>")+7:xml.find("</entry>")]
             xml = xml[xml.find("</entry>")+8:]
@@ -238,8 +283,8 @@ def process_authors_str(authors):
 def process_authors_list(string):
     """input a string of authors, return a list of authors"""
     authors = []
-    list_auth = string.split("and").strip()
+    list_auth = string.split("and")
     for author in list_auth:
         if author != "et al.":
-            authors.append(author)
+            authors.append(author.strip())
     return authors