Hello :) Today is Day 226!
A quick summary of today:
- finishing and documenting the personal finance assistant project
The full information and code is on my github, and below I will give some info about today’s improvements.
I added a prepare_data.py script
For reproducability, I prepared a script that takes the raw dataset, preprocesses it, indexes, creates description - all files needed for the UI to work, and outputs it to the folder where they can be directly read by the app.
Added a Makefile for easy setup
Usage: make [option]
Options:
help Show this help message
install Install dependencies
data Prepare data for the app (will overwrite any existing data)
start Start app
del-audio Delete any generated audio files
The last option is because the google TTS creates 3 mp3 files (for the welcome msg, user audio input and AI output), so for easy cleanup - all are deleted.
Switched to a different streamlit mic recorder
I was using another 3rd party mic input module, but it did not look very well in the UI and it did not work sometimes.
Old:
It had all these buttons and while recording, a big audio box appears.
The new one
A simple button that even changes text from - Start Audio to Stop Audio (and I can edit it). Bottom line - it provides better user experience.
Added prompt validity check
If the user inputs something that does not seem logical, or related to transactions, or even ‘…a u h’ (random letters) or if the transcribed audio is random letters because of some noise, the input first goes through gpt4o-mini to check whether:
- it is related to transactions
- related to the chat history
- not random letters
Maybe there is a better way to do it, but at the moment the code is:
def check_valid_prompt(user_text_query: str, chat_history: List[Dict[str, str]]) -> str:
"""
Check if the user query is valid and related to finance transactions.
"""
check_prompt = f"You are a personal finance assistant. Here is the conversation history so far: {chat_history}\n\nDoes the below user query look correct and related to finance transactions and the chat history? Answer with 'Yes' or 'No':\n\n{user_text_query}"
gpt4omini = gpt_client()
response = gpt4omini.chat.completions.create(
model="gpt-4o-mini",
messages=[{"role": "user", "content": check_prompt}]
)
llm_check_response = response.choices[0].message.content
return llm_check_response
Added project info
- Used Dataset - Kaggle
- Database - SQLite
- LLM - OpenAI’s GPT
- Audio Transcription Model - openai/whisper-small from HuggingFace
- RAG pipeline - LlamaIndex
- Monitoring/Tracing - Azire Pheonix
- User Interface - Streamlit
UI example usage:
Monitoring example:
Created a project diagram:
Recorder a short demo
Unfortunately, the demo is best seen in the repo itself.
That is all for today!
See you tomorrow :)