feat: upgrade to claude 3 and gpt 4o

app/basic_langchain/chat_models.py

@@ -54,7 +54,7 @@ def __call__(self, messages: List[AbstractMessage]) -> AIMessage:
             # claude wants system messages as a kwarg
             system = messages[0].content
             messages.pop(0)
-        response = self.client.beta.messages.create(
+        response = self.client.messages.create(
             model=self.model,
             system=system,
             temperature=self.temperature,

app/requirements.txt

@@ -1,5 +1,5 @@
 langchain[llms]==0.1.2
-anthropic~=0.15.1
+anthropic~=0.25.1
 stanza~=1.5.0
 openai~=1.12.0
 typer~=0.4.1

app/summarize_commentary/summarize_commentary.py

@@ -44,7 +44,7 @@ def summarize_commentary(source: TopicPromptSource, topic: Topic, company='opena
         print(f"Number of commentary tokens: {num_tokens}")
         completion = get_completion_openai(prompt)
     elif company == 'anthropic':
-        llm = ChatAnthropic(model="claude-instant-1.2", temperature=0)
+        llm = ChatAnthropic(model="claude-3-haiku-20240307", temperature=0)
         completion = llm([HumanMessage(content=prompt)]).content
     else:
         raise Exception("No valid company passed. Options are 'openai' or 'anthropic'.")

app/topic_modelling/poc.py

@@ -78,7 +78,7 @@ def get_raw_topics(text, lang):
     user_prompt = PromptTemplate.from_template("# Input\n{text}")
     human_message = HumanMessage(content=user_prompt.format(text=text))
 
-    # llm = ChatOpenAI(model="gpt-4", temperature=0)
+    # llm = ChatOpenAI(model="gpt-4o", temperature=0)
     llm = ChatAnthropic(model="claude-2", temperature=0)
 
     response = llm([system_message, human_message])

app/topic_prompt/contextualize.py

@@ -33,7 +33,7 @@ def context_from_surrounding_text(source: TopicPromptSource) -> str:
     human_message = HumanMessage(content=f"<segment>{segment_text}</segment>\n"
                                          f"<context>{section_text}</context>\n"
                                          f"<hint>{source.context_hint}</hint>")
-    llm = ChatAnthropic(model="claude-2.1", temperature=0)
+    llm = ChatAnthropic(model="gpt-4o", temperature=0)
     response = llm([system_message, human_message])
     context = get_by_xml_tag(response.content, "relevant_context").strip()
     if context is None:
@@ -44,7 +44,7 @@ def context_from_surrounding_text(source: TopicPromptSource) -> str:
 
 def general_context(source: TopicPromptSource):
     text = get_source_text_with_fallback(source, "en", auto_translate=True)
-    llm = ChatOpenAI(model="gpt-4", temperature=0)
+    llm = ChatOpenAI(model="gpt-4o", temperature=0)
     system_message = SystemMessage(content=f"""
     Given a text from the Jewish cannon, add any relevant context that would help a user understand this text from a
     Jewish perspective. Relevant context may be:

app/topic_prompt/differentiate_writing.py

@@ -13,7 +13,7 @@
 
 
 def differentiate_writing(sentence, phrase_to_avoid):
-    llm = ChatOpenAI(model="gpt-4", temperature=0)
+    llm = ChatOpenAI(model="gpt-4o", temperature=0)
     system_message = SystemMessage(content="Given a target sentence (wrapped in <target> tags) and a phrase to avoid "
                                            "using (wrapped in <phrase_to_avoid> tags), rewrite <target> so it doesn't "
                                            "use <phrase_to_avoid> but maintains the same meaning, tone and terminology."
@@ -30,7 +30,7 @@ def differentiate_writing(sentence, phrase_to_avoid):
 
 
 def remove_dependent_clause(sentence, phrase_to_avoid):
-    llm = ChatOpenAI(model="gpt-4", temperature=0)
+    llm = ChatOpenAI(model="gpt-4o", temperature=0)
     system_message = SystemMessage(content="Given a target sentence (wrapped in <target> tags) and a phrase to avoid "
                                            "using (wrapped in <phrase_to_avoid> tags):\n"
                                            "1) Determine if <phrase_to_avoid> includes an explanation of a term\n"
@@ -55,7 +55,7 @@ def remove_dependent_clause(sentence, phrase_to_avoid):
 
 
 def repeated_phrase(sentence, comparison_sentences: list[str]):
-    llm = ChatOpenAI(model="gpt-4", temperature=0)
+    llm = ChatOpenAI(model="gpt-4o", temperature=0)
     system_message = SystemMessage(content="Given a target sentence (wrapped in <target> tags) and a list of "
                                            "comparison sentences (each wrapped in <comparison> tags):\n"
                                            "1) Decide if there is a phrase shared between the target sentence and at least one comparison sentence\n"

app/topic_prompt/topic_prompt_generator.py

@@ -17,7 +17,7 @@ def _get_toprompt_options(lang: str, topic: Topic, source: TopicPromptSource, ot
                           num_tries=1, phrase_to_avoid=None) -> TopromptOptions:
     # TODO pull out formatting from _get_input_prompt_details
     llm_prompt = TopromptLLMPrompt(lang, topic, source, other_sources).get()
-    llm = ChatOpenAI(model="gpt-4", temperature=0)
+    llm = ChatOpenAI(model="gpt-4o", temperature=0)
     human_message = HumanMessage(content=llm_prompt.format())
     responses = []
     topic_prompts = []
@@ -67,7 +67,7 @@ def _improve_title(curr_responses, curr_title):
                                                        f"Rewrite the title, rephrasing to avoid using a colon."
                                                        f" Wrap the title in <title> tags. It should at most"
                                                        f" five words and grab the reader's attention.")
-    llm = ChatOpenAI(model="gpt-4", temperature=0.5)
+    llm = ChatOpenAI(model="gpt-4o", temperature=0.5)
     title_response = llm(curr_responses + [HumanMessage(content=better_title_prompt.format())])
     title_match = re.search(r'<title>(.+?)</title>', title_response.content)
     if title_match is None:

app/topic_prompt/uniqueness_of_source.py

@@ -40,7 +40,7 @@ def get_uniqueness_of_source(source: TopicPromptSource, topic: Topic, other_sour
 
 def summarize_based_on_uniqueness(text: str, uniqueness: str) -> str:
 
-    llm = ChatOpenAI(model="gpt-4", temperature=0)
+    llm = ChatOpenAI(model="gpt-4o", temperature=0)
     system_message = SystemMessage(content=
                                    "You are an intelligent Jewish scholar who is knowledgeable in all aspects of the Torah and Jewish texts.\n"
                                    "# Task\n"
@@ -90,7 +90,7 @@ def _get_uniqueness_of_source_as_compared_to_other_sources(source: TopicPromptSo
                                           '"hint": "{hint}"'
                                           '}}}}')
     human_message = HumanMessage(content=prompt.format(**prompt_inputs))
-    llm = ChatOpenAI(model="gpt-4", temperature=0)
+    llm = ChatOpenAI(model="gpt-4o", temperature=0)
 
     response = llm([system_message, human_message])
     uniqueness = re.sub(fr'^"?{uniqueness_preamble}\s*', '', response.content)

app/util/openai.py

@@ -2,13 +2,13 @@
 import tiktoken
 
 
-def count_tokens_openai(prompt, model="gpt-4"):
+def count_tokens_openai(prompt, model="gpt-4o"):
     encoding = tiktoken.encoding_for_model(model)
     num_tokens = len(encoding.encode(prompt))
     return num_tokens
 
 
-def get_completion_openai(prompt, model="gpt-4"):
+def get_completion_openai(prompt, model="gpt-4o"):
     messages = [{"role": "user", "content": prompt}]
     response = openai.ChatCompletion.create(
         model=model,