latest update of app.py, newer version of langchain package in requirements.txt, and tesla_db Vector database with 448 embedding dimensions and overlap of 24

app.py

@@ -1,10 +1,17 @@
+# +++
+# Import the libraries
+#---------------------------------------------------------------------------------------------------------
+# Import libraries for issuing OS commands.  In addition to the built-in format using the '!' scape character prefix
 import os
+# Tokenizing and data formatting
 import uuid
 import json
+# Data management
+import numpy as np
+import pandas as pd
+# For File path operations
 from pathlib import Path
 
-# ATTENTION: some versions in the requirements.txt file are more current than the ones used in the notebook
-
 # GUI components
 import gradio as gr
 
@@ -14,6 +21,7 @@ import gradio as gr
 from openai import OpenAI, OpenAIError
 
 # Embedding operations & Vector DB creation
+# from langchain_core.documents import Document
 from langchain_community.embeddings.sentence_transformer import SentenceTransformerEmbeddings
 from langchain_community.vectorstores import Chroma
 
@@ -23,13 +31,17 @@ from huggingface_hub import CommitScheduler
 # Set working directory (HuggingFace space)
 hf_space_dir = os.getcwd()
 
+# Obtain current directory and data file path
+hf_space_app_dir_path = Path.cwd()
+print(f"HuggingFace Space application directory: {hf_space_app_dir_path}\n")
+
 # Anyscale model hosting platform NOT USED in this notebook.  Open AI model hosting platform used instead.
 # client = OpenAI(
 #     base_url="https://api.endpoints.anyscale.com/v1",
 #     api_key=os.environ['ANYSCALE_API_KEY']
 # )
 
-#----------------------------------------------------------------------
+#--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
 # OpenAI API key stored as a "secret" HuggingFace Space
 # OPENAI_API_KEY 
 
@@ -63,8 +75,11 @@ client = OpenAI(
     # api_key=os.environ.get("OPENAI_API_KEY"),
     api_key=openai_api_key,
 )
-#---------------------------------------------------------------------
 
+print(f"OpenAI client created and authenticated with API key.\nUsing OpenAI model: {model_name}\n")
+#-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+
+print(f"Loading Vector DB from HuggingFace Space file space...\n")
 # embedding_model = SentenceTransformerEmbeddings(model_name='thenlper/gte-small')  
 # The gte-small model from OpenAI's family of models, which includes the GTE models designed for retrieval tasks, uses a specific number of embedding dimensions. The gte-small model has 384 embedding dimensions.
 # This dimensionality allows the model to capture semantic information effectively while maintaining a relatively small model size for efficiency in retrieval tasks.
@@ -77,27 +92,46 @@ embedding_model = SentenceTransformerEmbeddings(model_name='thenlper/gte-large')
 tesla_10k_collection = 'tesla-10k-2019-to-2023'
 
 # Example: Creating a collection with the correct dimensionality
-# tesla_10k_collection = Chroma.create_collection("tesla-10k-2019-to-2023", embedding_dim=384)
+# tesla_10k_collection = Chroma.create_collection("tesla-10k-2019-to-2023", embedding_dim=1024)
+
 
+persisted_vectordb_path = Path.joinpath(hf_space_app_dir_path, 'tesla_db')   # this is a pathlib object
 
+# persisted_vectordb_location = persisted_vectordb_path   # this is a pathlib object  ... this produces error in the Chroma parameter "persist_directory", as it is expecting a string object, and not a pathlib object.
+persisted_vectordb_location = str(persisted_vectordb_path)   # convert path to string
+
+print(f"Vector database location:\n{persisted_vectordb_location}\n")
 
 # vector database constructor Chroma()
 vectorstore_persisted = Chroma(
-    collection_name=tesla_10k_collection,
-    persist_directory='./tesla_db',
-    embedding_function=embedding_model
+    collection_name = tesla_10k_collection,
+    persist_directory = persisted_vectordb_location     # './tesla_db',
+    embedding_function = embedding_model
 )
 
+# Return VectorStoreRetriever initialized from this VectorStore.
 retriever = vectorstore_persisted.as_retriever(
-    search_type='similarity',
-    search_kwargs={'k': 5}
+    search_type = 'similarity',
+    search_kwargs = {'k': 5}
 )
+# Args:
+#     search_type (Optional[str]): Defines the type of search that the Retriever should perform.
+# Can be "similarity" (default), "mmr", or "similarity_score_threshold".
+#     search_kwargs (Optional[Dict]): Keyword arguments to pass to the
+#         search function. Can include things like:
+#             k: Amount of documents to return (Default: 4)
+#             score_threshold: Minimum relevance threshold for similarity_score_threshold
+
+print(f"Successfully obtained VectorStoreRetriever initialized from the Vector database.\n")
 
-# Prepare the logging functionality
 
+# Prepare the logging functionality
 log_file = Path("logs/") / f"data_{uuid.uuid4()}.json"
 log_folder = log_file.parent
+print(f"\nLogging dataset information:\n\tlog_file: {log_file}\n\tlog_folder: {log_folder}\n")
+
 
+# Scheduler will log every 2 API calls:
 scheduler = CommitScheduler(
     repo_id="document-qna-chroma-openai-logs",      # name of the log folder containing json elements -->  HuggingFace dataset       # OLD name: "document-qna-chroma-anyscale-logs",
     repo_type="dataset",
@@ -106,60 +140,115 @@ scheduler = CommitScheduler(
     every=2                                         # execute every two API calls
 )
 
+print(f"Retrieval Augmented Generation (RAG) Q&A\nLLM Prompt initialization... (System prompt and user_input template)\n")
+
+# LLM System Prompt
 qna_system_message = """
 You are an assistant to a financial services firm who answers user queries on annual reports.
-Users will ask questions delimited by triple backticks, that is, ```.
 User input will have the context required by you to answer user questions.
 This context will begin with the token: ###Context.
 The context contains references to specific portions of a document relevant to the user query.
-Please answer only using the context provided in the input. However, do not mention anything about the context in your answer. 
+
+User questions will begin with the token: ###Question, and the question text will be delimited by triple backticks, that is, ```.
+
+Please answer only using the context provided in the input. Do not mention anything about the context in your final answer.
+
 If the answer is not found in the context, respond "I don't know".
 """
 
+# LLM user_input template
 qna_user_message_template = """
 ###Context
-Here are some documents that are relevant to the question.
+Here are some documents that are relevant to the question mentioned below.
 {context}
+
+###Question
 ```
 {question}
 ```
 """
 
-# Define the predict function that runs when 'Submit' is clicked or when a API request is made
+# ANOTHER VERSION:
+# # LLM System Prompt
+# qna_system_message = """
+# You are an assistant to a financial services firm who answers user queries on annual reports.
+# Users will ask questions delimited by triple backticks, that is, ```.
+# User input will have the context required by you to answer user questions.
+# This context will begin with the token: ###Context.
+# The context contains references to specific portions of a document relevant to the user query.
+# Please answer only using the context provided in the input. However, do not mention anything about the context in your answer. 
+# If the answer is not found in the context, respond "I don't know".
+# """
+
+# # LLM user_input template
+# qna_user_message_template = """
+# ###Context
+# Here are some documents that are relevant to the question.
+# {context}
+# ```
+# {question}
+# ```
+# """
+
+# Define the "predict function" which will take the user_input, obtain the relevant context to answer the user question more accurately, and pass
+# both to the OpenAI client to make predictions using the OpenAI LLM model
+# The function runs when 'Submit' is clicked or when a API request is made
+#-------------------------------------------------------------------------------------------------------------------------------------------------------------
 def predict(user_input):
 
-    relevant_document_chunks = retriever.invoke(user_input)
-    # relevant_document_chunks = retriever.get_relevant_documents(query = user_input)
-    context_list = [doc_chunk.page_content for doc_chunk in relevant_document_chunks]
-    context_for_query = ". ".join(context_list)
+    # COMPOSING THE RESPONSE
+
+    # Retrieving relevant documents
+    relevant_document_chunks = retriever.get_relevant_documents(query = user_input)      # relevant_document_chunks = retriever.invoke(user_input)
+    print(f"Relevant document chunks = {len(relevant_document_chunks)}")
+    print(f"RELEVANT DOCUMENT CHUNKS TO BE USED AS CONTEXT TO ANSWER THE USER QUESTION:\n")
+    print("-"*80)
+    i = 0
+    for document in relevant_document_chunks:
+        print(f"\nDocument chunk {i+1}:")
+        i += 1
+        print(f"Metadata:\nSource: {document.metadata['source']}\nPage: {document.metadata['page']}\n")
+        print(f"Page content:\n-------------")
+        print(document.page_content.replace('\t', ' '))   # replace all tabs used as separators by default with a single space
+        print("-"*80)
     
+    context_list = [doc_chunk.page_content for doc_chunk in relevant_document_chunks]   # doc_chunk.page_content.replace('\t', ' ')   # replace all tabs used as separators by default with a single space
+
+    context_for_query = ". ".join(context_list)
+    # (method)
+    # join(__iterable: Iterable[LiteralString], /) -> LiteralString
+    # join(__iterable: Iterable[str], /) -> str
+
     prompt = [
         {'role':'system', 'content': qna_system_message},
-        {'role':'user', 'content': qna_user_message_template.format(
-            context=context_for_query,
-            question=user_input
+        {'role': 'user', 'content': qna_user_message_template.format(
+            context = context_for_query,
+            question = user_input
             )
         }
     ]
 
     try:
         response = client.chat.completions.create(
-            model=model_name,                # previous model used: 'mlabonne/NeuralHermes-2.5-Mistral-7B',
+            model=model_name,               # previous model used: 'mlabonne/NeuralHermes-2.5-Mistral-7B',
             messages=prompt,
-            temperature=0,                   # Temperature > 0 to encourage creative answer... Temperature = 0.7: A common setting that provides a balance between creativity and coherence.
-            # max_tokens=200                 # Limit the number of tokens in the response 
+            temperature=0,                  # Temperature > 0 to encourage creative answer... Temperature = 0.7: A common setting that provides a balance between creativity and coherence.
+            # max_tokens=400                # Limit the number of tokens in the response
         )
-
         prediction = response.choices[0].message.content.strip()   # Access response attributes directly
-
+    # Handle API errors
+    except openai.OpenAIError as e:
+        prediction = f'Sorry, I encountered the following OpenAI error: \n {e}'
     except Exception as e:
-        prediction = e
+        prediction = f'Sorry, I encountered the following error: \n {e}'
+
 
     # While the prediction is made, log both the inputs and outputs to a local log file (i.e., HuggingFace dataset)
     # While writing to the log file, ensure that the commit scheduler is locked to avoid parallel
-    # access (i.e., put a lock on the state of the log_file in case user are entering queries while the log operation is in progress.)
+    # access (i.e., put a lock on the state of the log_file in case users are entering queries while the log operation is in progress.)
+    # Write user_input, context and prediction to a HuggingFace dataset repo for logging
+    # Each time we get a prediction we will determine if we should log it to a hugging_face dataset according to the scheduler definition outside this function
     # Note: the log_file is a json file.
-    
     with scheduler.lock:
         with log_file.open("a") as f:
             # json.dumps turns the dictionary into a json string containing 'user_input', 'context_for_query', and 'prediction'
@@ -171,19 +260,31 @@ def predict(user_input):
                 }
             ))
             f.write("\n")     # write a new line to prepare for the next observation to be logged
-    
-    return prediction
+
+    prediction_result = prediction
+    print(f"\nPrediction result: {prediction_result} - {type(prediction_result)}\n")
+
+    return (prediction_result)
+#-------------------------------------------------------------------------------------------------------------------------------------------------------------
 
 
-textbox = gr.Textbox(placeholder="Enter your query here", lines=6)
+# Set up UI components for input and output
+# Input components
+user_question_textbox = gr.Textbox(placeholder="Enter your query here", lines=6)
+# Output components
+model_prediction = gr.Label(label="Model prediction")
+# model_prediction = "text"
 
 # Create the interface
 demo = gr.Interface(
-    inputs=textbox, fn=predict, outputs="text",
-    title="Ask Me Anything (AMA) on Tesla 10-K statements",
-    description="This web API presents an interface to ask questions on contents of the Tesla 10-K reports for the period 2019 - 2023.",
-    article="Note that questions that are not relevant to the Tesla 10-K report will not be answered.",
-    concurrency_limit=16
+    fn = predict, 
+    inputs = user_question_textbox, 
+    outputs = model_prediction    # "text",
+    title = "Ask Me Anything (AMA) on Tesla 10-K statements",
+    description= " This web API presents an interface to ask questions about the contents of the Tesla 10-K reports for the period 2019 - 2023.",
+    article = "Note that questions that are not relevant to the Tesla 10-K report will not be answered.",
+    allow_flagging="auto",    # automatically push to the HuggingFace Dataset
+    concurrency_limit = 16
 )
 
 
@@ -200,5 +301,7 @@ demo = gr.Interface(
 #     concurrency_limit=16
 # )
 
+# Launch container hosted by HuggingFace with a load balancer
 demo.queue()
-demo.launch()
+demo.launch(share=False)
+# To create a public link, set "share=True" in launch() ....  but if I execute this app.py locally, then I have to have my computer on for the public users to access the browser interface

requirements.txt

@@ -1,5 +1,8 @@
 openai==1.23.2
 chromadb==0.4.22
-langchain==0.1.1
-langchain-community==0.0.13
-sentence-transformers==2.3.1
+langchain==0.1.9
+langchain-community==0.0.32
+sentence-transformers==2.3.1
+pathlib==1.0.1
+pandas==2.1.4
+numpy==1.26.4

tesla_db/chroma.sqlite3

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:06ed87d12e6ee1b0b1f1f81fdfa4576e32f095c86fb7a2a66d0200a7640da76a
-size 46223360
+oid sha256:16fbd2946c0267a17248a40998bc5a3446a65dee1b8b5b356b6018da86cabbf0
+size 47091712

tesla_db/e8d4536e-a48d-4048-ad1d-9aaa0bb6ec69/data_level0.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:250d230dee83b006de1787c35948d598ddcffe236f81eafabbe31805e7469749
+size 12708000

tesla_db/e8d4536e-a48d-4048-ad1d-9aaa0bb6ec69/header.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e07096989a6d97db01f79643d6aa43690488b6a066eb20e594135825e0e34a70
+size 100

tesla_db/e8d4536e-a48d-4048-ad1d-9aaa0bb6ec69/index_metadata.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbc8823be5a3c3b78663ff93c9d05602df74d072b371471e88126f8df52eaabd
+size 172004

tesla_db/e8d4536e-a48d-4048-ad1d-9aaa0bb6ec69/length.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe6a12fdf9e9d1787634155c2f404ffea281dfd121ddd597deb71e17317f2576
+size 12000

tesla_db/e8d4536e-a48d-4048-ad1d-9aaa0bb6ec69/link_lists.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6322a3a45d75cc9f31cc0eafac6458a325d4c8e51525be50cc57030505294c41
+size 25736