merging from main

.github/workflows/sync_2_hf.yaml

@@ -17,4 +17,4 @@ jobs:
       - name: Push to hub
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
-        run: git push https://sabazo:[email protected]/spaces/sabazo/insurance-advisor-agents main
+        run: git push https://sabazo:[email protected]/spaces/sabazo/insurance_advisor_wb main

.gitignore

@@ -166,4 +166,6 @@ cython_debug/
 *.pickle
 
 
+# Databases
+
 *.db

README.md

@@ -3,7 +3,7 @@ title: Insurance Advisor Agents PoC
 emoji: 🤖
 colorFrom: red
 colorTo: indigo
-sdk: docker
+sdk: gradio
 python: 3.11
 app_file: app_gui.py
 pinned: false

app_gui.py

@@ -1,6 +1,7 @@
 # Import Gradio for UI, along with other necessary libraries
 import gradio as gr
 from rag_app.loading_data.load_S3_vector_stores import get_chroma_vs
+from rag_app.loading_data.load_S3_vector_stores import get_chroma_vs
 from rag_app.agents.react_agent import agent_executor
 from config import db
 

example.env

@@ -5,14 +5,12 @@ GOOGLE_API_KEY=""
 
 # Vectorstore storage on S3 and locally
 S3_LOCATION="rad-rag-demos"
-#FAISS_VS_NAME="vectorstores/faiss-insurance-agent-mpnet-1500.zip"
-#FAISS_VS_NAME="vectorstores/faiss-insurance-agent-MiniLM-L12-1500.zip"
 FAISS_VS_NAME="vectorstores/faiss-insurance-agent-multilingual-cased-1500.zip"
-CHROMA_VS_NAME="vectorstores/chroma-insurance-agent-multilingual-cased-1500.zip"
-#CHROMA_VS_NAME="vectorstores/chroma-insurance-agent-mpnet-1500.zip"
-#CHROMA_VS_NAME="vectorstore/chroma-insurance-agent-MiniLM-L12-1500.zip"
+CHROMA_VS_NAME="vectorstores/chroma-zurich-mpnet-1500.zip"
+# directories that need to be adjusted for windows
 FAISS_INDEX_PATH = "./vectorstore/faiss-insurance-agent-multilingual-cased-1500"
 CHROMA_DIRECTORY = "./vectorstore/chroma-insurance-agent-multilingual-cased-500"
+VS_DESTINATION_FOLDER="./vectorstore/"
 
 # for chromadb
 VECTOR_DATABASE_LOCATION="./vectorstore/chroma-insurance-agent-multilingual-cased-500"

rag_app/agents/react_agent.py

@@ -7,17 +7,17 @@ from langchain.agents.output_parsers import ReActJsonSingleInputOutputParser
 from langchain.tools.render import render_text_description
 import os
 from dotenv import load_dotenv
-from rag_app.structured_tools.agent_tools import (
-    web_research, ask_user, get_email
+# local cache
+from langchain.globals import set_llm_cache
+from langchain.cache import SQLiteCache # sqlite
+#from langchain.cache import InMemoryCache # in memory cache
+from rag_app.structured_tools.structured_tools import (
+    google_search, knowledgeBase_search
 )
 
 from langchain.prompts import PromptTemplate
 from rag_app.templates.react_json_with_memory_ger import template_system
 # from innovation_pathfinder_ai.utils import logger
-# from langchain.globals import set_llm_cache
-# from langchain.cache import SQLiteCache
-
-# set_llm_cache(SQLiteCache(database_path=".cache.db"))
 # logger = logger.get_console_logger("hf_mixtral_agent")
 
 config = load_dotenv(".env")
@@ -25,10 +25,8 @@ HUGGINGFACEHUB_API_TOKEN = os.getenv('HUGGINGFACEHUB_API_TOKEN')
 GOOGLE_CSE_ID = os.getenv('GOOGLE_CSE_ID')
 GOOGLE_API_KEY = os.getenv('GOOGLE_API_KEY')
 LLM_MODEL = os.getenv('LLM_MODEL')
-# LANGCHAIN_TRACING_V2 = "true"
-# LANGCHAIN_ENDPOINT = "https://api.smith.langchain.com"
-# LANGCHAIN_API_KEY = os.getenv('LANGCHAIN_API_KEY')
-# LANGCHAIN_PROJECT = os.getenv('LANGCHAIN_PROJECT')
+
+set_llm_cache(SQLiteCache(database_path=".cache.db"))
 
 # Load the model from the Hugging Face Hub
 llm = HuggingFaceEndpoint(repo_id=LLM_MODEL, 
@@ -40,11 +38,10 @@ llm = HuggingFaceEndpoint(repo_id=LLM_MODEL,
 
 
 tools = [
-    #knowledgeBase_search,
-    #google_search,
-    web_research,
-    ask_user,
-    get_email
+    knowledgeBase_search,
+    google_search,
+    #web_research,
+    #ask_user
     ]
 
 prompt = PromptTemplate.from_template(
@@ -74,8 +71,8 @@ agent_executor = AgentExecutor(
     agent=agent, 
     tools=tools, 
     verbose=True,
-    max_iterations=10,       # cap number of iterations
-    #max_execution_time=60,  # timout at 60 sec
+    max_iterations=20,       # cap number of iterations
+    max_execution_time=90,  # timout at 60 sec
     return_intermediate_steps=True,
     handle_parsing_errors=True,
     )

rag_app/chains/__init__.py

@@ -1,2 +1 @@
-# from rag_app.chains.s
 from rag_app.chains.user_response_sentiment_chain import user_response_sentiment_prompt

rag_app/loading_data/load_S3_vector_stores.py

@@ -32,41 +32,43 @@ embeddings = SentenceTransformerEmbeddings(model_name=model_name)
 
 ## FAISS
 def get_faiss_vs():
-    # Initialize an S3 client with unsigned configuration for public access
-    s3 = boto3.client('s3', config=Config(signature_version=UNSIGNED))
+    if not os.path.exists(FAISS_INDEX_PATH):
+        # Initialize an S3 client with unsigned configuration for public access
+        s3 = boto3.client('s3', config=Config(signature_version=UNSIGNED))
 
-    # Define the destination for the downloaded file
-    VS_DESTINATION = FAISS_INDEX_PATH + ".zip"
-    try:
-        # Download the pre-prepared vectorized index from the S3 bucket
-        print("Downloading the pre-prepared FAISS vectorized index from S3...")
-        s3.download_file(S3_LOCATION, FAISS_VS_NAME, VS_DESTINATION)
+        # Define the destination for the downloaded file
+        VS_DESTINATION = FAISS_INDEX_PATH + ".zip"
+        try:
+            # Download the pre-prepared vectorized index from the S3 bucket
+            print("Downloading the pre-prepared FAISS vectorized index from S3...")
+            s3.download_file(S3_LOCATION, FAISS_VS_NAME, VS_DESTINATION)
 
-        # Extract the downloaded zip file
-        with zipfile.ZipFile(VS_DESTINATION, 'r') as zip_ref:
-            zip_ref.extractall('./vectorstore/')
-        print("Download and extraction completed.")
-        return FAISS.load_local(FAISS_INDEX_PATH, embeddings, allow_dangerous_deserialization=True)
-        
-    except Exception as e:
-        print(f"Error during downloading or extracting from S3: {e}", file=sys.stderr)
-    #faissdb = FAISS.load_local(FAISS_INDEX_PATH, embeddings)
+            # Extract the downloaded zip file
+            with zipfile.ZipFile(VS_DESTINATION, 'r') as zip_ref:
+                zip_ref.extractall('./vectorstore/')
+            print("Download and extraction completed.")
+            return FAISS.load_local(FAISS_INDEX_PATH, embeddings, allow_dangerous_deserialization=True)
+            
+        except Exception as e:
+            print(f"Error during downloading or extracting from S3: {e}", file=sys.stderr)
+        #faissdb = FAISS.load_local(FAISS_INDEX_PATH, embeddings)
 
 
 ## Chroma DB
 def get_chroma_vs():
-    # Initialize an S3 client with unsigned configuration for public access
-    s3 = boto3.client('s3', config=Config(signature_version=UNSIGNED))
+    if not os.path.exists(CHROMA_DIRECTORY):
+        # Initialize an S3 client with unsigned configuration for public access
+        s3 = boto3.client('s3', config=Config(signature_version=UNSIGNED))
 
-    VS_DESTINATION = CHROMA_DIRECTORY+".zip"
-    try:
-        # Download the pre-prepared vectorized index from the S3 bucket
-        print("Downloading the pre-prepared chroma vectorstore from S3...")
-        s3.download_file(S3_LOCATION, CHROMA_VS_NAME, VS_DESTINATION)
-        with zipfile.ZipFile(VS_DESTINATION, 'r') as zip_ref:
-            zip_ref.extractall('./vectorstore/')
-        print("Download and extraction completed.")
-        chromadb = Chroma(persist_directory=CHROMA_DIRECTORY, embedding_function=embeddings)
-        #chromadb.get()
-    except Exception as e:
-        print(f"Error during downloading or extracting from S3: {e}", file=sys.stderr)
+        VS_DESTINATION = CHROMA_DIRECTORY+".zip"
+        try:
+            # Download the pre-prepared vectorized index from the S3 bucket
+            print("Downloading the pre-prepared chroma vectorstore from S3...")
+            s3.download_file(S3_LOCATION, CHROMA_VS_NAME, VS_DESTINATION)
+            with zipfile.ZipFile(VS_DESTINATION, 'r') as zip_ref:
+                zip_ref.extractall('./vectorstore/')
+            print("Download and extraction completed.")
+            chromadb = Chroma(persist_directory=CHROMA_DIRECTORY, embedding_function=embeddings)
+            #chromadb.get()
+        except Exception as e:
+            print(f"Error during downloading or extracting from S3: {e}", file=sys.stderr)

rag_app/loading_data/load_chroma_db_cross_platform.py

@@ -0,0 +1,59 @@
+from pathlib import Path
+import boto3
+from botocore.client import Config
+from botocore import UNSIGNED
+from dotenv import load_dotenv
+import os
+import sys
+import zipfile
+
+
+def download_chroma_from_s3(s3_location:str,
+                            chroma_vs_name:str,
+                            vectorstore_folder:str,
+                            vs_save_name:str) -> None:
+    """
+    Downloads the Chroma DB from an S3 storage to local folder
+        
+        Args
+            s3_location (str): The name of S3 bucket
+            chroma_vs_name (str): The name of the file to download from S3
+            vectorstore_folder (str): The filepath to vectorstore folder in project dir
+            vs_save_name (str): The name of the vector store
+
+    """
+    vs_destination = Path()/vectorstore_folder/vs_save_name
+    vs_save_path = vs_destination.with_suffix('.zip')
+
+    try:
+        # Initialize an S3 client with unsigned configuration for public access
+        s3 = boto3.client('s3', config=Config(signature_version=UNSIGNED))
+        s3.download_file(s3_location, chroma_vs_name, vs_save_path)
+        print('Downloaded file from S3')
+
+        # Extract the zip file
+        with zipfile.ZipFile(file=str(vs_save_path), mode='r') as zip_ref:
+            zip_ref.extractall(path=vectorstore_folder)
+        print("Extracted zip file")
+
+    except Exception as e:
+        print(f"Error during downloading or extracting from S3: {e}", file=sys.stderr)
+
+    # Delete the zip file
+    vs_save_path.unlink()
+    print("Deleting zip file")
+
+if __name__ == "__main__":
+
+    S3_LOCATION = os.getenv("S3_LOCATION")
+
+    chroma_vs_name = "vectorstores/chroma-zurich-mpnet-1500.zip"
+    
+    project_dir = Path().cwd().parent.parent
+    vs_destination = str(project_dir / 'vectorstore')
+    assert Path(vs_destination).is_dir(), "Cannot find vectorstore folder"
+
+    download_chroma_from_s3(s3_location=S3_LOCATION, 
+                            chroma_vs_name=chroma_vs_name, 
+                            vectorstore_folder=vs_destination,
+                            vs_save_name='chroma-zurich-mpnet-1500')

rag_app/reranking.py

@@ -5,11 +5,13 @@ from dotenv import load_dotenv
 import os
 from langchain_community.embeddings import HuggingFaceInferenceAPIEmbeddings
 import requests
+from langchain_community.vectorstores import Chroma
+
 
 load_dotenv()
 
 
-def get_reranked_docs(query:str, 
+def get_reranked_docs_faiss(query:str, 
                       path_to_db:str, 
                       embedding_model:str,
                       hf_api_key:str, 
@@ -59,22 +61,71 @@ def get_reranked_docs(query:str,
         ranked_results = sorted(zip(docs, passages, relevance_scores), key=lambda x: x[2], reverse=True)
         top_k_results = ranked_results[:num_docs]
         return [doc for doc, _, _ in top_k_results]
+    
 
- 
-if __name__ == "__main__":
+
+def get_reranked_docs_chroma(query:str, 
+                      path_to_db:str, 
+                      embedding_model:str,
+                      hf_api_key:str,
+                      reranking_hf_url:str = "https://api-inference.huggingface.co/models/sentence-transformers/all-mpnet-base-v2", 
+                      num_docs:int=5) -> list:
+    """ Re-ranks the similarity search results and returns top-k highest ranked docs
+
+        Args:
+            query (str): The search query
+            path_to_db (str): Path to the vectorstore database
+            embedding_model (str): Embedding model used in the vector store
+            num_docs (int): Number of documents to return
+        
+        Returns: A list of documents with the highest rank
+    """
+    embeddings = HuggingFaceInferenceAPIEmbeddings(api_key=hf_api_key,
+                                                   model_name=embedding_model)
+    # Load the vectorstore database
+    db = Chroma(persist_directory=path_to_db, embedding_function=embeddings)
     
-    HUGGINGFACEHUB_API_TOKEN = os.getenv('HUGGINGFACEHUB_API_TOKEN')
-    EMBEDDING_MODEL = os.getenv("EMBEDDING_MODEL")
+    # Get k documents based on similarity search
+    sim_docs =  db.similarity_search(query=query, k=10)
+
+    passages = [doc.page_content for doc in sim_docs]
     
-    path_to_vector_db = Path("..")/'vectorstore/faiss-insurance-agent-500'
+    # Prepare the payload
+    payload = {"inputs": 
+               {"source_sentence": query,
+	            "sentences": passages}}
+    
+    headers = {"Authorization": f"Bearer {hf_api_key}"}
 
-    query = "Ich möchte wissen, ob ich meine geriatrische Haustier-Eidechse versichern kann"
+    response = requests.post(url=reranking_hf_url, headers=headers, json=payload)
+    print(f'{response = }')
+    if response.status_code != 200:
+        print('Something went wrong with the response')
+        return
     
-    top_5_docs = get_reranked_docs(query=query,
-                                   path_to_db=path_to_vector_db,
-                                   embedding_model=EMBEDDING_MODEL,
-                                   hf_api_key=HUGGINGFACEHUB_API_TOKEN,
-                                   num_docs=5)
+    similarity_scores = response.json()
+    ranked_results = sorted(zip(sim_docs, passages, similarity_scores), key=lambda x: x[2], reverse=True)
+    top_k_results = ranked_results[:num_docs]
+    return [doc for doc, _, _ in top_k_results]
     
-    for i, doc in enumerate(top_5_docs):
-        print(f"{i}: {doc}\n")
+
+
+if __name__ == "__main__":
+
+   
+    HUGGINGFACEHUB_API_TOKEN = os.getenv('HUGGINGFACEHUB_API_TOKEN')
+    EMBEDDING_MODEL = "sentence-transformers/multi-qa-mpnet-base-dot-v1"
+
+    project_dir = Path().cwd().parent
+    path_to_vector_db = str(project_dir/'vectorstore/chroma-zurich-mpnet-1500')
+    assert Path(path_to_vector_db).exists(), "Cannot access path_to_vector_db "
+
+    query = "I'm looking for student insurance"
+
+    re_ranked_docs = get_reranked_docs_chroma(query=query,
+                                              path_to_db= path_to_vector_db,
+                                              embedding_model=EMBEDDING_MODEL,
+                                              hf_api_key=HUGGINGFACEHUB_API_TOKEN)
+
+
+    print(f"{re_ranked_docs=}")

rag_app/structured_tools/structured_tools.py

@@ -4,6 +4,10 @@ from langchain_community.embeddings.sentence_transformer import (
     SentenceTransformerEmbeddings,
 )
 from langchain_community.vectorstores import Chroma
+import ast
+from rag_app.loading_data.load_S3_vector_stores import get_chroma_vs
+import chromadb
+
 from rag_app.utils.utils import (
     parse_list_to_dicts, format_search_results
 )
@@ -11,6 +15,10 @@ import chromadb
 import os
 from config import db, PERSIST_DIRECTORY, EMBEDDING_MODEL
 
+persist_directory = os.getenv('VECTOR_DATABASE_LOCATION')
+embedding_model = os.getenv("EMBEDDING_MODEL")
+if not os.path.exists(persist_directory):
+    get_chroma_vs()
 
 @tool
 def memory_search(query:str) -> str:

rag_app/templates/react_json_with_memory_ger.py

@@ -1,7 +1,7 @@
 template_system = """
-Du bist ein freundlicher Versicherungsproduktberater. Deine Aufgabe ist es, Kunden dabei zu helfen, die besten Produkte der Württembergische GmbH zu finden\
-und ihnen mehr informationen dazu per Email zusenden, wenn du seine Fragen beanwortest hast.\
+Du bist ein freundlicher Versicherungsproduktberater. Deine Aufgabe ist es, Kunden dabei zu helfen, die besten Produkte der Württembergische GmbH zu finden.\
 Hilfe dem Benutzer, Antworten auf seine Fragen zu finden. Antworte kurz und einfach und biete an, dem Benutzer das Produkt und die Bedingungen zu erklären.\
+Wenn du denkst, die Fragen des Benutzers ausreichend beantowrtet zu haben, Frage ihn nach seiner Email Addresse.\
 Beantworte die folgenden Fragen so gut du kannst. Du hast Zugriff auf die folgenden Tools:
 
 <TOOLS>