Update models/langOpen.py

Updated prompt. Changed temp to 0.01. Added compression retriever.

models/langOpen.py

@@ -9,18 +9,17 @@ from langchain.embeddings import OpenAIEmbeddings
 from langchain.prompts import PromptTemplate
 from langchain_pinecone import PineconeVectorStore
 
-prompt_template = """Answer the question using the given context to the best of your ability. 
-If you don't know, answer I don't know.
+prompt_template = """You are an expert on California Drinking Water Regulations.
+Answer the question solely using relevant regulations in the given context. DO NOT USE ANY OTHER SOURCES.
+If the given context does not contain the relevant information, say so.
 Context: {context}
 Topic: {topic}
 
 Use the following example format for your answer:
-[FORMAT]
 Answer:
     The answer to the user question.
 Reference:
     The list of references to the specific sections of the documents that support your answer.    
-[END_FORMAT]
 """
 
 
@@ -30,7 +29,7 @@ PROMPT = PromptTemplate(template=prompt_template, input_variables=["context", "t
 class LangOpen:
     def __init__(self, model_name: str) -> None:
         self.index = self.initialize_index("langOpen")
-        self.llm = ChatOpenAI(temperature=0.3, model=model_name)
+        self.llm = ChatOpenAI(temperature=0.01, model=model_name)
         self.chain = LLMChain(llm=self.llm, prompt=PROMPT)
 
     def initialize_index(self, index_name):
@@ -43,7 +42,13 @@ class LangOpen:
     def get_response(self, query_str):
         print("query_str: ", query_str)
         print("model_name: ", self.llm.model_name)
-        docs = self.index.similarity_search(query_str, k=4)
+        #docs = self.index.similarity_search(query_str, k=4)
+        vectorstore_retriever = self.index.as_retriever(search_type="similarity", search_kwargs={"k": 10})
+        compressor = CohereRerank()
+        compression_retriever = ContextualCompressionRetriever(
+        base_compressor=compressor, base_retriever=vectorstore_retriever
+        )
+        docs = compression_retriever.get_relevant_documents(query_str)
         inputs = [{"context": doc.page_content, "topic": query_str} for doc in docs]
         result = self.chain.apply(inputs)[0]["text"]
         return result