Skip to main content

​
Overview

RAG Chat provides four core functions for managing document processing, vector storage, and question answering. All functions are designed to work seamlessly with LangChain and OpenAI embeddings.

​
load_existing_vector_store()

Loads an existing Chroma vector store from the persistent directory.

​
Signature

def load_existing_vector_store()

​
Returns

​
vector_store
Chroma | None
Returns a Chroma vector store instance if the persistent directory exists, otherwise returns None.

​
Behavior

  • Checks if the db directory exists in the project root
  • If found, initializes a Chroma vector store with OpenAI embeddings
  • Uses the same embedding function as document ingestion for consistency

​
Example

from app import load_existing_vector_store

# Load existing vector store on app startup
vector_store = load_existing_vector_store()

if vector_store:
    print("Vector store loaded successfully")
else:
    print("No existing vector store found")

​
Source Reference

Defined in app.py:16-23

​
process_file(file)

Processes a PDF file and splits it into document chunks suitable for vector storage.

​
Signature

def process_file(file)

​
Parameters

​
file
UploadedFile
required
A file object from Streamlit’s file uploader. Must be a PDF file.

​
Returns

​
chunks
list[Document]
A list of LangChain Document objects, split into chunks with overlap for better context preservation.

​
Processing Details

  1. Temporary File Creation: Creates a temporary file with .pdf suffix
  2. PDF Loading: Uses PyPDFLoader to extract text from all pages
  3. Text Splitting: Applies RecursiveCharacterTextSplitter with:
    • chunk_size: 1000 characters
    • chunk_overlap: 400 characters (40% overlap for context preservation)
  4. Cleanup: Automatically removes the temporary file after processing

​
Example

from app import process_file

# In a Streamlit app
uploaded_file = st.file_uploader("Upload PDF", type="pdf")

if uploaded_file:
    chunks = process_file(uploaded_file)
    print(f"Processed {len(chunks)} document chunks")

​
Source Reference

Defined in app.py:26-40

​
add_to_vector_store(documents, vector_store=None)

Adds document chunks to an existing vector store or creates a new one.

​
Signature

def add_to_vector_store(documents, vector_store=None)

​
Parameters

​
documents
list[Document]
required
A list of LangChain Document objects to add to the vector store. Typically output from process_file().
​
vector_store
Chroma
default:"None"
An existing Chroma vector store instance. If None, a new vector store will be created.

​
Returns

​
vector_store
Chroma
The updated or newly created Chroma vector store instance.

​
Behavior

  • If vector_store is provided: Adds documents to the existing store using add_documents()
  • If vector_store is None: Creates a new Chroma vector store with:
    • OpenAI embeddings
    • Persistent storage in the db directory

​
Example

from app import load_existing_vector_store, process_file, add_to_vector_store

# Load existing store or start fresh
vector_store = load_existing_vector_store()

# Process and add new documents
for uploaded_file in uploaded_files:
    chunks = process_file(uploaded_file)
    vector_store = add_to_vector_store(chunks, vector_store)

print("All documents added to vector store")

​
Source Reference

Defined in app.py:43-52

​
ask_question(model, query, vector_store)

Processes a user question using RAG (Retrieval-Augmented Generation) with conversation history.

​
Signature

def ask_question(model, query, vector_store)

​
Parameters

​
model
str
required
The OpenAI model identifier to use for generating responses. Supported models:
  • gpt-3.5-turbo
  • gpt-4
  • gpt-4-turbo
  • gpt-4o-mini
  • gpt-4o
​
query
str
required
The user’s question or prompt.
​
vector_store
Chroma
required
A Chroma vector store instance containing the document embeddings for retrieval.

​
Returns

​
response
str
The AI-generated response in markdown format, based on retrieved context and conversation history.

​
RAG Pipeline

  1. LLM Initialization: Creates a ChatOpenAI instance with the specified model
  2. Retriever Setup: Converts vector store to a retriever for semantic search
  3. Prompt Construction:
    • System prompt instructs the model to answer based on context
    • Includes full conversation history from st.session_state.messages
    • Responses are formatted in markdown with interactive visualizations
  4. Chain Execution:
    • Retrieves relevant document chunks
    • Passes context and query to the LLM
    • Returns formatted response

​
System Prompt Behavior

The function uses a Portuguese system prompt that:
  • Instructs the model to use retrieved context for answers
  • Asks for explanations when information is not available
  • Requests markdown formatting with interactive visualizations

​
Example

from app import ask_question

# Assuming vector_store exists and st.session_state.messages is initialized
response = ask_question(
    model="gpt-4o-mini",
    query="What are the main findings in the uploaded documents?",
    vector_store=vector_store
)

print(response)  # Markdown-formatted answer

​
Source Reference

Defined in app.py:55-82

​
Notes

All functions require the OPENAI_API_KEY environment variable to be set. See Environment Variables for setup instructions.
The process_file() function only supports PDF files. Other file formats will cause errors. See Supported Formats for details.

Build docs developers (and LLMs) love

Get started for freeTalk to us