Ask your IRIS classes with Ollama, IRIS VectorDB and Langchain

If you want to know if a class about a topic already exists asking a simple natural language question, it is possible now. Download and run the application https://openexchange.intersystems.com/package/langchain-iris-tool to know all about your project classes in a Chat.

Installation:

$ git clone https://github.com/yurimarx/langchain-iris-tool.git
$ docker-compose build
$ docker-compose up -d

Using:

1. Open the URL http://localhost:8501

2. Check out the Settings button used to the Agent connect the InterSystems IRIS

3. Ask about your developed classes (e.g.: Are there classes that inherit from Persistent?)

Solutions used:

1. Ollama - private LLM and NLP Chat tool
2. Lanchain - plataform to build AI agents
3. Streamlit - Frontend framework
4. InterSystems IRIS as a server to answer the questions about it

About Ollama

It is a free and on-premises LLM solution to be able running Generative AI with privacy and security because your data will be processed on-premises only. The project Ollama supports many models including mistral, Open AI models, Deepseek models and others running on-premises. This package used Ollama via docker compose with the model mistral:

ollama:
    image: ollama/ollama:latest
    deploy:
      resources:
        reservations:
          devices:
          - driver: nvidia
            capabilities: ["gpu"]
            count: all  # Adjust count for the number of GPUs you want to use
    ports:
      - 11434:11434
    volumes:
      - ./model_files:/model_files
      - .:/code
      - ./ollama:/root/.ollama
    container_name: ollama_iris
    pull_policy: always
    tty: true
    entrypoint: ["/bin/sh", "/model_files/run_ollama.sh"] # Loading the finetuned Mistral with the GGUF file
    restart: always
    environment:
      - OLLAMA_KEEP_ALIVE=24h
      - OLLAMA_HOST=0.0.0.0

About Langchain:

Langchain it is a framework to build GenAI applications easily. The Langchain has the concept of tool. Tools are plug-ins (RAG applications) used by Langchain to complement the work of the LLMs. This application implemented a langchain tool to ask management and development questions for your IRIS server:

from langchain.llms import Ollama
from langchain.chains import RetrievalQA
from langchain.document_loaders import CSVLoader
from langchain.embeddings import OllamaEmbeddings
from langchain_iris import IRISVector

def get_insights(question, csv_file, iris_conn, collection_name):
    
    # Load and process the CSV data    
    loader = CSVLoader(csv_file)
    documents = loader.load()

    llm = Ollama(
        base_url="http://ollama:11434", 
        model="mistral", 
        temperature=0,
    )

    # Create embeddings
    embeddings = OllamaEmbeddings(model="mistral", base_url="http://ollama:11434", temperature=0)

    db = IRISVector.from_documents(
        embedding=embeddings, 
        documents=documents,
        connection_string=iris_conn,
        collection_name=collection_name,
        pre_delete_collection=True
    )

    qa = RetrievalQA.from_chain_type(
        llm=llm,
        chain_type="stuff",
        retriever=db.as_retriever())
    
    return qa({"query": question})

About Streamlit:

The streamlit solution it is used to develop frontends using python language. This application has a streamlit chat application to interact with the Ollama, Langchain and IRIS and get relevant responses:

import pandas as pd
import streamlit as st
from sqlalchemy import create_engine
import langchain_helper as lch

username = "_system"
password = "SYS"
hostname = "iris"
port = 51972
webport = 52773
namespace = "USER"
st.set_page_config(page_title="InterSystems IRIS Classes Demo", page_icon="📜")

st.title("Langchain IRIS Classes Chat")

with st.popover("Settings"):
    with st.spinner(text="Connecting to the IRIS classes"):
        engine = create_engine("iris://" + username + ":" + password + "@" + hostname + ":" + str(port) + "/" + namespace)
        connection  = engine.connect()
        query = 'select * from %Dictionary.ClassDefinition where substring(ID,1,1) <> \'%\' and  Copyright is null'
        df = pd.read_sql(query, con=connection)
        df.to_csv("classes.csv")
    
    username = st.text_input("Username:", username)
    password = st.text_input("Password:", password)
    hostname = st.text_input("Hostname:", hostname)
    port = int(st.text_input("Port:", port))
    webport = int(st.text_input("Web port:", webport))
    namespace = st.text_input("Namespace:", namespace)

            

# User query input
query = st.text_input(label="Enter your query")

# Submit button
if st.button(label="Ask IRIS Classes", type="primary"):
    
    with st.spinner(text="Generating response"):
        iris_conn_str = f"iris://{username}:{password}@{hostname}:{port}/{namespace}"
        response = lch.get_insights(query, "classes.csv", iris_conn=iris_conn_str, collection_name="classes")
        st.write(response['result'])