import streamlit as st from azure.storage.blob import BlobServiceClient from azure.core.exceptions import ResourceExistsError from azure.keyvault.secrets import SecretClient from azure.identity import DefaultAzureCredential import chromadb from chromadb.config import Settings import openai from openai.embeddings_utils import get_embedding from app_utils import create_prompt, list_sources

Connect to key vault

credential = DefaultAzureCredential() key_vault_name = "microhack-key-vault" key_vault_uri = f"https://{key_vault_name}.vault.azure.net" client = SecretClient(vault_url=key_vault_uri, credential=credential)

Blob storage

account_name = "microhack" account_key = client.get_secret("BLOB-KEY").value container_name = "documents"

blob_service_client = BlobServiceClient( account_url=f"https://{account_name}.blob.core.windows.net", credential=account_key ) container_client = blob_service_client.get_container_client(container_name)

OpenAI

openai_api_key = client.get_secret("OPENAI-KEY").value openai_endpoint = client.get_secret("OPENAI-ENDPOINT").value

openai.api_type = "azure" openai.api_key = openai_api_key openai.api_base = openai_endpoint openai.api_version = "2022-12-01"

Chroma

chroma_address = client.get_secret("CHROMA-DB-ADDRESS").value

chroma_client = chromadb.Client( Settings( chroma_api_impl="rest", chroma_server_host=chroma_address, chroma_server_http_port="8000", ) )

Get collection

collection = chroma_client.get_collection("microhack-collection")

App title

st.title("Microhack: Semantic Q&A-Bot")

File upload in sidebar (only allows PDFs)

doc = st.sidebar.file_uploader( ":page_facing_up: Upload your own documents to the knowledge base here", type=["pdf"], )

Upload doc to blob storage and display status banner

if doc is not None: blob_client = container_client.get_blob_client(doc.name)

try:
    blob_client.upload_blob(doc)
    st.sidebar.success("Document uploaded successfully!", icon="🚀")
except ResourceExistsError:
    st.sidebar.error("Document was already uploaded!", icon="🛑")

Query free-text window

st.markdown("### ❓ Query the bot here:") query = st.text_input("query", value="Who are statworx?", label_visibility="collapsed")

Paragraph slider

n_paragraphs = st.slider( "Number of paragraphs to retrieve and source answer from:", min_value=1, max_value=5, value=3, )

Embed query

query_embedding = get_embedding(query, engine="microhack-ada-text-embedding")

Query chroma collection

response = collection.query( query_embeddings=[query_embedding], n_results=n_paragraphs, include=["documents"] )

Generate answer with completion endpoint

completions = openai.Completion.create( engine="microhack-davinci-003-text-completion", # the deployed model temperature=0.3, # level of creativity in the response prompt=create_prompt( # the retrieved paragraphs + query + fixed instructions response=response, # the retrieved paragraphs query=query, # the query ), max_tokens=n_paragraphs * 125, # maximum tokens in both the prompt and completion, scales with n_paragraphs n=1, # number of generated answers )

Display answer

st.markdown("### :robot_face: The Q&A-bot answers:") st.write(completions["choices"][0]["text"])

Display paragraphs

st.markdown(f"### 💡 This answer was sourced from {n_paragraphs} paragraphs:") st.write(list_sources(response))