open ai update to gpt4

models/custom_parsers.py

@@ -20,38 +20,38 @@ class CustomStringOutputParser(BaseOutputParser[List[str]]):
         text_list = [x.strip() for x in list(chain.from_iterable(text_list))]
         return text_list
     
-class CustomINSTOutputParser(BaseOutputParser[List[str]]):
-    """Parse the output of an LLM call to a list."""
+# class CustomINSTOutputParser(BaseOutputParser[List[str]]):
+#     """Parse the output of an LLM call to a list."""
 
-    name = "Kit"
-    name_rx = re.compile(r""+ name + r":|" + name.lower() + r":")
-    whispers = re.compile((r"([\(]).*?([\)])"))
-    reactions = re.compile(r"([\*]).*?([\*])")
-    double_spaces = re.compile(r"  ")
-    quotation_rx = re.compile('"')
+#     name = "Kit"
+#     name_rx = re.compile(r""+ name + r":|" + name.lower() + r":")
+#     whispers = re.compile((r"([\(]).*?([\)])"))
+#     reactions = re.compile(r"([\*]).*?([\*])")
+#     double_spaces = re.compile(r"  ")
+#     quotation_rx = re.compile('"')
 
-    @property
-    def _type(self) -> str:
-        return "str"
+#     @property
+#     def _type(self) -> str:
+#         return "str"
 
-    def parse_whispers(self, text: str) -> str:
-        text = self.name_rx.sub("", text).strip()
-        text = self.reactions.sub("", text).strip()
-        text = self.whispers.sub("", text).strip()
-        text = self.double_spaces.sub(r" ", text).strip()
-        text = self.quotation_rx.sub("", text).strip()
-        return text 
+    # def parse_whispers(self, text: str) -> str:
+    #     text = self.name_rx.sub("", text).strip()
+    #     text = self.reactions.sub("", text).strip()
+    #     text = self.whispers.sub("", text).strip()
+    #     text = self.double_spaces.sub(r" ", text).strip()
+    #     text = self.quotation_rx.sub("", text).strip()
+    #     return text 
     
-    def parse_split(self, text: str) -> str:
-        text = text.split("[INST]")[0]
-        text_list = text.split("[/INST]")
-        text_list = [x.split("</s>") for x in text_list]
-        text_list = [x.strip() for x in list(chain.from_iterable(text_list))]
-        text_list = [x.split("\n\n") for x in text_list]
-        text_list = [x.strip().rstrip("\n") for x in list(chain.from_iterable(text_list))]
-        return text_list
+    # def parse_split(self, text: str) -> str:
+    #     text = text.split("[INST]")[0]
+    #     text_list = text.split("[/INST]")
+    #     text_list = [x.split("</s>") for x in text_list]
+    #     text_list = [x.strip() for x in list(chain.from_iterable(text_list))]
+    #     text_list = [x.split("\n\n") for x in text_list]
+    #     text_list = [x.strip().rstrip("\n") for x in list(chain.from_iterable(text_list))]
+    #     return text_list
     
-    def parse(self, text: str) -> str:
-        text = self.parse_whispers(text)
-        text_list = self.parse_split(text)
-        return text_list
+    # def parse(self, text: str) -> str:
+    #     text = self.parse_whispers(text)
+    #     text_list = self.parse_split(text)
+    #     return text_list

models/openai/role_models.py

@@ -1,88 +1,45 @@
 import logging
-import pandas as pd
 from models.custom_parsers import CustomStringOutputParser
-from utils.app_utils import get_random_name
 from langchain.chains import ConversationChain
-from langchain.llms import OpenAI
+from langchain_openai import ChatOpenAI
 from langchain.prompts import PromptTemplate
+from models.business_logic_utils.input_processing import initialize_conversation
 
-
-ISSUE_MAPPING = {
-    "anxiety": "issue_Anxiety",
-    "suicide": "issue_Suicide",
-    "safety_planning": "issue_Suicide",
-    "GCT": "issue_Gral",
-}
-
-EN_TEXTER_TEMPLATE_ = """The following is a conversation between you and a crisis counselor.
-{current_issue}
-You are able to reply with what the character should say. You are able to reply with your character's dialogue inside and nothing else. Do not write explanations.
-Do not disclose your name unless asked.
-Current conversation:
-{history}
-helper: {input}
+OPENAI_TEMPLATE = """{template}
+{{history}}
+helper: {{input}}
 texter:"""
 
-SP_TEXTER_TEMPLATE_ = """La siguiente es una conversacion contigo y un consejero de crisis
-{current_issue}
-Puedes responder como lo haria tu personaje. Puedes responder como si fueras tu personaje y nada mas. No escribas explicaciones
-No reveles tu nombre a menos que te lo pregunten
-Conversacion Actual:
-{history}
-helper: {input}
-texter:"""
-
-CURRENT_ISSUE_MAPPING = {
-    "issue_Suicide-en": "Your character, {texter_name}, has suicidal thoughts. Your character has a plan to end his life and has all the means and requirements to do so. {seed}",
-    "issue_Anxiety-en": "Your character, {texter_name}, is experiencing anxiety. Your character has suicide thoughts but no plan. {seed}",
-    "issue_Suicide-es": "Tu personaje, {texter_name}, tiene pensamientos suicidas. Tu personaje tiene un plan para terminar con su vida y tiene todos los medios y requerimientos para hacerlo. {seed}",
-    "issue_Anxiety-es": "Tu personaje, {texter_name}, experimenta ansiedad. Tu personaje tiene pensamientos suicidas pero ningun plan. {seed}",
-    "issue_Gral-en": "Your character {texter_name} is experiencing a mental health crisis. {seed}",
-    "issue_Gral-es": "Tu personaje {texter_name} esta experimentando una crisis de salud mental. {seed}",
-}
-
-def get_template_role_models(issue: str, language: str, texter_name: str = "", seed="") -> str:
-    """_summary_
-
-    Args:
-        issue (str): Issue for template, current options are ['issue_Suicide','issue_Anxiety']
-        language (str): Language for the template, current options are ['en','es']
-        texter_name (str): texter to apply to template, defaults to None
-
-    Returns:
-        str: template
-    """
-    current_issue = CURRENT_ISSUE_MAPPING.get(
-        f"{issue}-{language}", CURRENT_ISSUE_MAPPING[f"issue_Gral-{language}"]
-    )
-    default_name = get_random_name()
-    current_issue = current_issue.format(
-        texter_name=default_name if not texter_name else texter_name,
-        seed = seed
-    )
-
-    if language == "en":
-        template = EN_TEXTER_TEMPLATE_.format(current_issue=current_issue, history="{history}", input="{input}")
-    elif language == "es":
-        template = SP_TEXTER_TEMPLATE_.format(current_issue=current_issue, history="{history}", input="{input}")
+def get_template_role_models(issue: str, language: str, texter_name: str = "") -> str:
+    model_input = {
+        "issue": issue,
+        "language": language,
+        "texter_name": texter_name,
+        "messages": [],
+    }
 
-    return template
+    # Initialize the conversation (adds the system message)
+    model_input = initialize_conversation(model_input, "")
+    return model_input["messages"][0]["content"]
 
 def get_role_chain(template, memory, temperature=0.8):
 
+    template = OPENAI_TEMPLATE.format(template=template)
     PROMPT = PromptTemplate(
         input_variables=['history', 'input'],
         template=template
     )
-    llm = OpenAI(
-        temperature=temperature, 
-        max_tokens=150,
+    llm = ChatOpenAI(
+        model="gpt-4o",
+        temperature=temperature,
+        max_tokens=256,
     )
     llm_chain = ConversationChain(
         llm=llm,
         prompt=PROMPT,
         memory=memory,
-        output_parser=CustomStringOutputParser()
+        output_parser=CustomStringOutputParser(),
+        verbose=True,
     )
-    logging.debug(f"loaded GPT3.5 model")
+    logging.debug(f"loaded GPT4o model")
     return llm_chain, "helper:"

requirements.txt

@@ -1,6 +1,5 @@
 scipy==1.11.1
-openai==1.7.0
-langchain==0.1.0
+langchain==0.3.0
 pymongo==4.5.0
 mlflow==2.9.0
-langchain-community==0.0.20
+langchain-openai==0.2.0

utils/chain_utils.py

@@ -12,8 +12,7 @@ def get_chain(issue, language, source, memory, temperature, texter_name=""):
         OA_engine = finetuned_models[f"{issue}-{language}"]
         return get_finetuned_chain(OA_engine, memory, temperature)
     elif source in ('OA_rolemodel'):
-        seed = seeds.get(issue, "GCT")['prompt']
-        template = get_template_role_models(issue, language, texter_name=texter_name, seed=seed)
+        template = get_template_role_models(issue, language, texter_name=texter_name)
         return get_role_chain(template, memory, temperature)
     elif source in ('CTL_llama2'):
         if language == "English":