Commit after using contextual compression

app.py

@@ -1,21 +1,17 @@
 from dotenv import load_dotenv
 load_dotenv() 
 
-from langchain_openai import ChatOpenAI
-from langchain_core.prompts import ChatPromptTemplate
-import tiktoken
-from langchain.text_splitter import RecursiveCharacterTextSplitter
-from langchain_community.document_loaders import PyMuPDFLoader
-from langchain_community.vectorstores import Qdrant
-from langchain_openai.embeddings import OpenAIEmbeddings
 from operator import itemgetter
-from langchain.schema.output_parser import StrOutputParser
 from langchain.schema.runnable import RunnablePassthrough
 from utils import *
 import os
 import getpass
 from langchain.globals import set_debug
 import chainlit as cl 
+from langchain_openai import ChatOpenAI, OpenAI
+from langchain.retrievers.multi_query import MultiQueryRetriever
+from langchain.retrievers import ContextualCompressionRetriever
+from langchain.retrievers.document_compressors import LLMChainExtractor
 
 
 class RAGMeta10K:
@@ -28,11 +24,25 @@ class RAGMeta10K:
         self.UtilsObject = Utils()
         self.rag_prompt_template = self.UtilsObject.init_prompt()
         self.UtilsObject.split_into_chunks()
+        # normail retriever
         self.qdrant_retriever = self.UtilsObject.get_vector_store().as_retriever()
+        
+        # MultiQuery Retriever
+        # self.mqr_retriever = MultiQueryRetriever.from_llm(
+        #     retriever=self.qdrant_retriever, llm=ChatOpenAI(temperature=0)
+        # )
+        
+        #Contexttual Compression
+        # gives correct answer for board question but messup on Cash question
+        compressor = LLMChainExtractor.from_llm(OpenAI(temperature=0))
+        self.compression_retriever = ContextualCompressionRetriever(
+            base_compressor=compressor, base_retriever=self.qdrant_retriever
+        )
+        
 
     def ask_question(self, question: str):
         retrieval_augmented_qa_chain = (
-            {"context": itemgetter("question") | self.qdrant_retriever, "question": itemgetter("question")}
+            {"context": itemgetter("question") | self.compression_retriever, "question": itemgetter("question")}
             | RunnablePassthrough.assign(context=itemgetter("context"))
             | {"response": self.rag_prompt_template | self.UtilsObject.get_llm_model(), "context": itemgetter("context")}
         )
@@ -55,6 +65,7 @@ class RAGMeta10K:
 @cl.on_chat_start
 async def start_chat():
     ragObject = RAGMeta10K()
+    # ragObject.UtilsObject.generate_test_set()
     cl.user_session.set("ragObject", ragObject)
 
 @cl.on_message

requirements.txt

@@ -6,4 +6,5 @@ langchain-community
 chainlit
 uvicorn
 pymupdf
-qdrant-client
+qdrant-client
+ragas

utils.py

@@ -4,9 +4,8 @@ from langchain_community.document_loaders import PyMuPDFLoader
 from langchain_core.prompts import ChatPromptTemplate
 from langchain.text_splitter import RecursiveCharacterTextSplitter
 from langchain_openai.embeddings import OpenAIEmbeddings
-from typing import List
-from langchain_core.documents.base import Document
 from langchain_community.vectorstores import Qdrant
+from ragas.testset.generator import TestsetGenerator
 
 
 class Utils:
@@ -15,19 +14,33 @@ class Utils:
         llm_name: str = "gpt-3.5-turbo",
         pdf_name: str = "meta-10k.pdf",
         embedding_model: str = "text-embedding-3-small",
+        generator_llm: str = "gpt-3.5-turbo-16k",
+        critic_llm: str= "gpt-4-turbo"
     ) -> None:
         self.openai_chat_model = ChatOpenAI(model=llm_name)
         self.enc = tiktoken.encoding_for_model(llm_name)
         self.docs = PyMuPDFLoader(pdf_name).load()
         self.embedding_model = OpenAIEmbeddings(model=embedding_model)
+        self.generator_llm = ChatOpenAI(model=generator_llm)
+        self.critic_llm = ChatOpenAI(model=critic_llm)
+        self.test_generator = TestsetGenerator.from_langchain(
+            generator_llm,
+            critic_llm,
+            OpenAIEmbeddings()
+        )
+
     
     #semanticTextSplitter
     #tokenRTextSplitter
     def split_into_chunks(self):
         text_splitter = RecursiveCharacterTextSplitter(
-            chunk_size=500, chunk_overlap=50, length_function=self.tiktoken_len
+            chunk_size=2000, chunk_overlap=200, length_function=self.tiktoken_len
         )
         self.split_chunks = text_splitter.split_documents(self.docs)
+        
+        # semantic splitter
+        # text_splitter = SemanticChunker(OpenAIEmbeddings())
+        # self.split_chunks = text_splitter.split_documents([self.docs])
         return self.split_chunks
 
     def get_llm_model(self):
@@ -61,3 +74,17 @@ class Utils:
             collection_name="meta-10k",
         )
         return self.qdrant_vectorstore
+    
+    def generate_test_set(self)-> None:
+        text_splitter_eval = RecursiveCharacterTextSplitter(
+            chunk_size = 600,
+            chunk_overlap = 50
+        )
+        eval_documents = text_splitter_eval.split_documents(self.docs)
+        distributions = {
+            "simple": 0.5,
+            "multi_context": 0.4,
+            "reasoning": 0.1
+        }
+        testset = self.test_generator.generate_with_langchain_docs(eval_documents, 20, distributions, is_async = False)
+        print("santhosh:"+len(testset.to_pandas()))