app20

app.py

@@ -51,8 +51,8 @@ loader = PyPDFDirectoryLoader('pdfs')
 data=loader.load()
 # split documents
 text_splitter = RecursiveCharacterTextSplitter(
-    chunk_size=500,
-    chunk_overlap=70,
+    chunk_size=1024,
+    chunk_overlap=150,
     length_function=len
 )
 docs = text_splitter.split_documents(data)
@@ -69,7 +69,7 @@ vectordb = Chroma.from_documents(
     persist_directory=persist_directory
 )
 # define retriever
-retriever = vectordb.as_retriever(search_kwargs={"k": 1}, search_type="mmr")
+retriever = vectordb.as_retriever(search_kwargs={"k": 2}, search_type="mmr")
 
 class FinalAnswer(BaseModel):
     question: str = Field(description="the original question")
@@ -81,7 +81,10 @@ parser = PydanticOutputParser(pydantic_object=FinalAnswer)
 template = """
 Your name is AngryGreta and you are a recycling chatbot with the objective to anwer questions from user in English or Spanish /
 Use the following pieces of context to answer the question /
-Answer in the same language of the question /
+If the question is English answer in English /
+If the question is Spanish answer in Spanish /
+Do not mention the word context when you answer a question, use the word database instead /
+Answer the question fully and provide as much relevant detail as possible. Do not cut your response short
 Context: {context}
 User: {question}
 {format_instructions}
@@ -99,10 +102,11 @@ llm = HuggingFaceHub(
     repo_id="mistralai/Mixtral-8x7B-Instruct-v0.1",
     task="text-generation",
     model_kwargs={
-        "max_new_tokens": 1024,
+        "max_new_tokens": 2000,
         "top_k": 30,
         "temperature": 0.1,
         "repetition_penalty": 1.03,
+        "early_stopping "="never"
     },
 )