initial commit

app.py

@@ -65,7 +65,7 @@ with gr.Blocks() as demo:
     with gr.Row():
         with gr.Column(scale=1, min_width=300):
             input_pdf = gr.File(label="Upload PDF", file_types=[".pdf"], file_count='single')
-            gr.Markdown("_Use a PDF to enhance the chatbot's knowledge!_", visible=not pdf_uploaded)
+            gr.Markdown("_Upload a PDF to chat with it!_", visible=not pdf_uploaded)
     
     with gr.Row():
         chatbot = gr.Chatbot(label="Chatbot Conversation", type="messages", bubble_full_width=True, show_copy_button=True, autoscroll=True)
@@ -84,4 +84,4 @@ with gr.Blocks() as demo:
     )
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch(server_port=9000,quiet=True,show_api=False,app_kwargs={"docs_url":'/docs'})

retriever.py

@@ -3,6 +3,9 @@ from langchain_openai import OpenAIEmbeddings
 from qdrant_client import QdrantClient
 from langchain_qdrant import QdrantVectorStore
 from qdrant_client.http import models
+from langchain_groq import ChatGroq
+from langchain_core.output_parsers import StrOutputParser
+from langchain_core.prompts import PromptTemplate
 
 from dotenv import load_dotenv
 
@@ -33,18 +36,64 @@ class Retriever():
                         'Indirect Tax Laws',
                         'INDIAN Income Tax ACTS',
                         'ONLINESITES']
+        self.groq = ChatGroq(model='llama3-70b-8192')
+
+    
+    
+    def multi_questions(self,user_prompt):
+        llm = self.groq
+        prompt = f'''
+# You are an excellent Query Decomposer for database retrieval optimization.
+# You are given a user_query.
+===============================
+# TASK:
+    -> Your task is to provide a structured and hierarchical breakdown of the user query.
+    -> This breakdown should be in the form of an ordered sequence that helps in extracting the right context from the database.
+    -> Build the user query from the bottom level (basic requirements) to the top level (more specific details), ensuring the retrieval context improves at each level.
+===============================
+# USER_QUERY: {{user}}
+===============================
+# EXAMPLE:
+    1. #USER_QUERY: "For 5 lakh, what type of taxes should I pay and how much?"
+       -> #EXPECTED OUTPUT: | I'm purchasing a car for 5 lakh. | What type of taxes should I pay on the purchase of automobiles? | What type of taxes should I pay on the purchase of a car for 5 lakh? |
+    
+    2. #USER_QUERY: "For 5 lakh, what type of taxes should I pay and how much?"
+       -> #EXPECTED OUTPUT: | NEW TAX REGIME and Income tax. | My income is 5 lakh. What type of taxes should I pay and how much should I pay? |
+
+===============================
+# OUTPUT FORMAT:
+    -> Provide the formatted output separated with the pipe '|' enclosed as: |...|...|
+    -> Stick to the given format without any additional explanation. Your only response must be the formatted sequence of queries.
+    -> Do not answer the user question directly. Your job is to provide the decomposed queries in the format shown in the examples.
+'''
+
+        rag_prompt = PromptTemplate.from_template(prompt)
+        l = (rag_prompt | llm | StrOutputParser())
+        stream = l.invoke({"user":user_prompt})
+        return stream
     
-    def filter(self,query):
+    def multiple_contexts(self,user_prompt):
+        questions = self.multi_questions(user_prompt).split("|")
+        contexts = []
+        num = 3
+        for i in questions[:num]:
+            if(i!='' and i!=' ' and i.strip()!=''):
+                contexts+=self.filter(i)
+            else:
+                num+=1
+        return contexts
+    
+    def filter(self,query,k1=7,k2=17):
         retriever1 = self.vector_store.as_retriever(
                                             search_type="similarity_score_threshold",
-                                            search_kwargs={"k": 7,
+                                            search_kwargs={"k": k1,
                                                            'score_threshold':0.7,
                                                             'filter':models.Filter(must=[models.FieldCondition(key="metadata.DOCUMENT_IS_ABOUT", match=models.MatchValue(value=self.filters[-1]),)])
                                                             },
                                         )
         retriever2 = self.vector_store.as_retriever(
                                             search_type="similarity_score_threshold",
-                                            search_kwargs={"k": 17,
+                                            search_kwargs={"k": k2,
                                                            'score_threshold':0.7,
                                                             'filter':models.Filter(must_not=[models.FieldCondition(key="metadata.DOCUMENT_IS_ABOUT", match=models.MatchValue(value=self.filters[-1]),)])
                                                            },
@@ -64,7 +113,8 @@ class Retriever():
         return ret
 
     def data_retrieve(self, query=''):
-        retrieved_docs = self.vector_store.similarity_search_with_score(query, k=20)
+        retrieved_docs = self.vector_store.similarity_search_with_score(query, k=5)
         return [doc for doc, _ in retrieved_docs]
 
-
+# ret = Retriever()
+# print(ret.multiple_contexts("i'm purchasing a car for 5Lack, what type of taxes should I pay and how much?"))

setup.py

@@ -49,32 +49,33 @@ class Script():
     
     def gpt_loaders(self,query:str,history:str):
         template= f"""
-                    # You are an excellent Question & Answering BOT. Given a question and the context you will answer the question only based on the given context.
+                    # You are an excellent Question & Answering BOT. Given a question and the context, you will answer the question only based on the given context.
                     # You will be given a user_query (or) User_question (or) User_scenario.
-                    # TASK: Your task is to provide an Answer to the USER_QUERY with the given CONTEXT_DATA.
                     ===============================
                     #USER_QUERY :  {{question}}
                     ===============================
-                    #METADATA_OF_CONTEXT : -> The context given is related to INDIAN-TAXATIONS.
-                                        -> It may contain how to calculate tax for GOODS/SERVICES/INDIVIDUAL/CARS/TRAINS/etc anything related to INDIAN TAXES.
-                                        -> Based on the user_query use the context accordingly.
-                                        -> You can also provide a rough calculation for an example if asked for tax calculations related from the CONTEXT (if it is available in the CONTEXT).
+                    #METADATA_OF_CONTEXT :
+                    -> The context given is related to INDIAN-TAXATION.
+                    -> It may contain how to calculate tax for GOODS/SERVICES/INDIVIDUAL/CARS/TRAINS/etc. — anything related to INDIAN TAXES.
+                    -> Consider providing information about tax types like GST, RTO tax, and additional charges where relevant.
                     #CONTEXT : {{context}}
                     ===============================
-                    You are also given previous ChatHistories (User question and corressponding AI answer) to you as an extra data.
-                    --# When to take the history as CONTEXT : Only if the history is relevant to the current question you are permitted to take the chat history as a context.
-                    --# If it is not relevant to the current question do not take it.
+                    You are also given previous ChatHistories (User question and corresponding AI answer) to you as extra data.
+                    --# When to take the history as CONTEXT: Only if the history is relevant to the current question, you are permitted to take the chat history as a context.
+                    --# If it is not relevant to the current question, do not take it.
                     #Chat History : {{history}}
                     ===============================
                     -> You are allowed to provide the answer only from the given context.
                     -> Don't provide your own answer that is not in the given context.
                     -> If you are not able to answer the given question from the context => PROVIDE "Sorry! Unable to find an answer for your question. Try Again."
-                    -> Try to be a precise and provide a proper output for the question. Don't explain any questions too lengthy max[100 words].
-                    -> Provide answer only to the question that is asked.
+                    -> Try to be precise and provide a proper output for the question. Don't explain any questions too lengthy max[100 words].
+                    -> Provide an answer only to the question that is asked.
                     ===============================
                     # OUTPUT FORMAT:
-                        -> Your output may be given to a voice model for a speech output. Try to be precise with your words. At the same time, fill the user with your answer
-                        -> Don't provide any etc explanation apart from the answer output.
+                    -> Your output may be given to a voice model for a speech output. Try to be precise with your words. At the same time, fill the user with your answer.
+                    -> Don't provide any further explanation apart from the answer output.
+                    -> Provide a brief breakdown of the different types of applicable taxes if relevant.
+                    -> Make sure to state the specific taxes for scenarios like vehicle purchases, real estate, etc., from the CONTEXT.
                 """
         rag_prompt = PromptTemplate.from_template(template)
         rag_chain = (
@@ -82,7 +83,7 @@ class Script():
                     | self.openai_client
                     | StrOutputParser()
                     )
-        question ={"context": self.format_docs(self.retriever.data_retrieve(query)), "question": query, "history": history}
+        question ={"context": self.format_docs(self.retriever.filter(query)), "question": query, "history": history}
         return rag_chain,question
     
     def gpt_loaders_id(self,query:str,history:str,id:str):