Upload 6 files

replaced langchain with open ai

app.py

@@ -1,21 +1,18 @@
 import gradio as gr
 from full_chain import get_response
 import os
-import openai
+import json
+from openai import OpenAI
 import json
 
-from langchain_openai import ChatOpenAI
-from langchain.schema import HumanMessage, SystemMessage
+# from langchain_openai import ChatOpenAI
+# from langchain.schema import HumanMessage, SystemMessage
 import urllib3
 
 urllib3.disable_warnings(urllib3.exceptions.InsecureRequestWarning)
-# api_key = os.getenv("OPENAI_API_KEY")
-# client = openai.OpenAI(api_key=api_key)
-# Initialize ChatOpenAI
-llm = ChatOpenAI(
-    model="gpt-4o-mini",
-    temperature=0
-)
+
+# Initialize OpenAI client
+client = OpenAI()  # It will automatically use OPENAI_API_KEY from environment
 
 def load_content(filename):
     """Load content from text files"""
@@ -34,54 +31,62 @@ def predict(message, history):
     """Process user message and return appropriate response."""
     try:
         # Query classification prompt
-        classifier_prompt = """You are the Tobacco Watcher Assistant. Analyze the user's query and categorize it into exactly ONE of these types:
+        classifier_messages = [
+            {"role": "system", "content": """You are the Tobacco Watcher Assistant. Analyze the user's query and categorize it into exactly ONE of these types:
 
-        1. HELP - Questions about using the website, its features, or navigation
-        Example: "How do I use filters?", "How to search for articles?"
+            1. HELP - Questions about using the website, its features, or navigation
+            Example: "How do I use filters?", "How to search for articles?"
 
-        2. ABOUT - Questions about Tobacco Watcher's purpose, mission, or organization
-        Example: "What is Tobacco Watcher?", "Who runs this website?"
+            2. ABOUT - Questions about Tobacco Watcher's purpose, mission, or organization
+            Example: "What is Tobacco Watcher?", "Who runs this website?"
 
-        3. FILTER - Requests for specific articles using filters
-        Example: "Show articles about smoking in India from 2023", "Find French articles about e-cigarettes"
+            3. FILTER - Requests for specific articles using filters
+            Example: "Show articles about smoking in India from 2023", "Find French articles about e-cigarettes"
 
-        4. QUERY - Questions seeking tobacco-related information
-        Example: "How many people smoke in Asia?", "What are the effects of secondhand smoke?"
+            4. QUERY - Questions seeking tobacco-related information
+            Example: "How many people smoke in Asia?", "What are the effects of secondhand smoke?"
 
-        Respond with ONLY the category name (HELP, ABOUT, FILTER, or QUERY).
-        """
-        
-        messages = [
-            SystemMessage(content=classifier_prompt),
-            HumanMessage(content=message)
+            Respond with ONLY the category name (HELP, ABOUT, FILTER, or QUERY)."""},
+            {"role": "user", "content": message}
         ]
         
-        response = llm.invoke(messages)
-        query_type = response.content.strip().upper()
+        completion = client.chat.completions.create(
+            model="gpt-4o-mini",
+            messages=classifier_messages,
+            temperature=0
+        )
+        query_type = completion.choices[0].message.content.strip().upper()
         print(f"Query type: {query_type}")
 
         if query_type == "HELP":
             help_content = load_content("help.txt")
-            messages = [
-                SystemMessage(content="""You are the Tobacco Watcher Help Assistant.
+            help_messages = [
+                {"role": "system", "content": """You are the Tobacco Watcher Help Assistant.
 Use the provided help content to guide users on how to use the platform's features.
-Be clear and specific in your instructions. If a feature isn't mentioned in the content, acknowledge that and suggest contacting support."""),
-                HumanMessage(content=f"Using this help content:\n\n{help_content}\n\nAnswer this question: {message}")
+Be clear and specific in your instructions. If a feature isn't mentioned in the content, acknowledge that and suggest contacting support."""},
+                {"role": "user", "content": f"Using this help content:\n\n{help_content}\n\nAnswer this question: {message}"}
             ]
-          
-            response = llm.invoke(messages)
-            return response.content
+            completion = client.chat.completions.create(
+                model="gpt-4o-mini",
+                messages=help_messages,
+                temperature=0
+            )
+            return completion.choices[0].message.content
 
         elif query_type == "ABOUT":
             about_content = load_content("about.txt")
-            messages = [
-                SystemMessage(content="""You are the Tobacco Watcher Assistant specializing in explaining the platform.
+            about_messages = [
+                {"role": "system", "content": """You are the Tobacco Watcher Assistant specializing in explaining the platform.
 Use the provided content to answer questions about Tobacco Watcher's purpose, mission, features, and organization.
-Be concise but informative. If a specific detail isn't in the content, say so rather than making assumptions."""),
-                HumanMessage(content=f"Using this content:\n\n{about_content}\n\nAnswer this question: {message}")
+Be concise but informative. If a specific detail isn't in the content, say so rather than making assumptions."""},
+                {"role": "user", "content": f"Using this content:\n\n{about_content}\n\nAnswer this question: {message}"}
             ]
-            response = llm.invoke(messages)
-            return response.content
+            completion = client.chat.completions.create(
+                model="gpt-4o-mini",
+                messages=about_messages,
+                temperature=0
+            )
+            return completion.choices[0].message.content
 
         elif query_type == "FILTER":
             filter_options = load_filter_options()
@@ -108,14 +113,18 @@ Be concise but informative. If a specific detail isn't in the content, say so ra
             
             url_prompt += "\nGenerate a valid URL for this query. Return ONLY the complete URL."
             
-            messages = [
-                SystemMessage(content=url_prompt),
-                HumanMessage(content=message)
+            url_messages = [
+                {"role": "system", "content": url_prompt},
+                {"role": "user", "content": message}
             ]
             
             try:
-                response = llm.invoke(messages)
-                url_response = response.content.strip()
+                completion = client.chat.completions.create(
+                    model="gpt-4o-mini",
+                    messages=url_messages,
+                    temperature=0
+                )
+                url_response = completion.choices[0].message.content.strip()
                 print(f"Generated URL: {url_response}")
                 
                 if url_response.startswith("http"):

feed_to_llm_v2.py

@@ -1,9 +1,4 @@
-from langchain_openai import ChatOpenAI
-
-from langchain.schema import (
-    HumanMessage,
-    SystemMessage
-)
+from openai import OpenAI
 import tiktoken
 import re
 
@@ -11,6 +6,7 @@ from get_articles import save_solr_articles_full
 from rerank import crossencoder_rerank_answer
 import logging
 from logging.handlers import RotatingFileHandler
+import os
 
 # Configure logging
 logger = logging.getLogger("TobaccoInfoAssistant")
@@ -23,11 +19,13 @@ formatter = logging.Formatter("%(asctime)s - %(levelname)s - %(message)s")
 handler.setFormatter(formatter)
 logger.addHandler(handler)
 
+# Initialize OpenAI client
+client = OpenAI()
+
 def num_tokens_from_string(string: str, encoder) -> int:
     num_tokens = len(encoder.encode(string))
     return num_tokens
 
-
 def feed_articles_to_gpt_with_links(information, question):
     prompt = """
     You are a Question Answering system specializing in tobacco-related topics. You have access to several curated articles, each numbered (e.g., Article 1, Article 2). These articles cover various aspects of tobacco use, health effects, legislation, and quitting resources.
@@ -60,6 +58,7 @@ def feed_articles_to_gpt_with_links(information, question):
     published_dates = [published_dates for score, contents, uuids, titles, domains, published_dates in information]
     logger.info(f"Article retrieved: {len(articles)}")
     logger.info(f"Article titles: {titles_list}")
+    
     for i in range(len(articles)):
         addition = f"Article {i + 1}: {articles[i]} {separator}"
         token_count += num_tokens_from_string(addition, encoder)
@@ -69,14 +68,18 @@ def feed_articles_to_gpt_with_links(information, question):
 
     prompt += content
     logger.info(f"Prompt: {prompt}")
-    llm = ChatOpenAI(model="gpt-4o-mini", temperature=0.0)
-    message = [
-        SystemMessage(content=prompt),
-        HumanMessage(content=question)
+    
+    messages = [
+        {"role": "system", "content": prompt},
+        {"role": "user", "content": question}
     ]
 
-    response = llm.invoke(message)
-    response_content = response.content  # Access the content of the AIMessage
+    completion = client.chat.completions.create(
+        model="gpt-4o-mini",
+        messages=messages,
+        temperature=0
+    )
+    response_content = completion.choices[0].message.content
     logger.info(f"LLM Response Content: {response_content}")
 
     # Extract sources from the response content
@@ -84,7 +87,7 @@ def feed_articles_to_gpt_with_links(information, question):
     parenthetical_matches = re.findall(r'\(Article \d+\)', response_content)
     
     if not (inline_matches or parenthetical_matches):
-        return response_content, [], [], []
+        return response_content, [], [], [], []
 
     # Combine and get unique article numbers
     all_matches = inline_matches + [m.strip('()') for m in parenthetical_matches]
@@ -122,7 +125,7 @@ def feed_articles_to_gpt_with_links(information, question):
     cited_published_dates = []
     for article_num in used_article_nums:
         uuid = uuids[article_num]
-        link = f"https://tobaccowatcher.globaltobaccocontrol.org/articles/{uuid}/"
+        link = f"https://tobaccowatcher.globaltobactocontrol.org/articles/{uuid}/"
         cited_links.append(link)
         cited_titles.append(titles_list[article_num])
         cited_domains.append(domains_list[article_num])
@@ -133,6 +136,8 @@ if __name__ == "__main__":
     question = "How is United States fighting against tobacco addiction?"
     rerank_type = "crossencoder"
     llm_type = "chat"
-    csv_path = save_solr_articles_full(question, keyword_type="rake")
+    from get_articles import save_solr_articles_full
+    from rerank import crossencoder_rerank_answer
+    csv_path = save_solr_articles_full(question, 15, keyword_type="rake")
     reranked_out = crossencoder_rerank_answer(csv_path, question)
     feed_articles_to_gpt_with_links(reranked_out, question)

requirements.txt

@@ -1,15 +1,9 @@
-gradio==4.25.0
-langchain==0.1.14
-langchain-core==0.1.40
-langchain-openai==0.1.1
-nltk==3.8.1
-openai==1.16.2
-pandas==2.2.1
-pysolr==3.9.0
-rake-nltk==1.0.6
-sentence-transformers==2.2.2
-tiktoken==0.5.2
-torch==2.1.2
-huggingface-hub==0.20.2
-python-dotenv==1.0.1
-docarray==0.40.0
+gradio==4.25.0
+openai
+nltk==3.8.1
+pandas==2.2.1
+pysolr==3.9.0
+rake-nltk==1.0.6
+sentence-transformers==2.2.2
+tiktoken==0.5.2
+python-dotenv==1.0.1