Create an Agentic RAG using OpenVINO and LlamaIndex

This Jupyter notebook can be launched after a local installation only.

An agent is an automated reasoning and decision engine. It takes in a user input/query and can make internal decisions for executing that query in order to return the correct result. The key agent components can include, but are not limited to:

• Breaking down a complex question into smaller ones

• Choosing an external Tool to use + coming up with parameters for calling the Tool

• Planning out a set of tasks

• Storing previously completed tasks in a memory module

LlamaIndex is a framework for building context-augmented generative AI applications with LLMs.LlamaIndex imposes no restriction on how you use LLMs. You can use LLMs as auto-complete, chatbots, semi-autonomous agents, and more. It just makes using them easier. You can build agents on top of your existing LlamaIndex RAG pipeline to empower it with automated decision capabilities. A lot of modules (routing, query transformations, and more) are already agentic in nature in that they use LLMs for decision making.

Agentic RAG = Agent-based RAG implementation

While standard RAG excels at simple queries across a few documents, agentic RAG takes it a step further and emerges as a potent solution for question answering. It introduces a layer of intelligence by employing AI agents. These agents act as autonomous decision-makers, analyzing initial findings and strategically selecting the most effective tools for further data retrieval. This multi-step reasoning capability empowers agentic RAG to tackle intricate research tasks, like summarizing, comparing information across multiple documents and even formulating follow-up questions -all in an orchestrated and efficient manner.

agentic-rag

This example will demonstrate using RAG engines as a tool in an agent with OpenVINO and LlamaIndex.

Table of contents:

• Prerequisites

• Download models

  • Download LLM

  • Download Embedding model

• Create models

  • Create OpenVINO LLM

  • Create OpenVINO Embedding

• Create tools

• Run Agentic RAG

Installation Instructions

This is a self-contained example that relies solely on its own code.

We recommend running the notebook in a virtual environment. You only need a Jupyter server to start. For details, please refer to Installation Guide.

Prerequisites

Install required dependencies

import os

os.environ["GIT_CLONE_PROTECTION_ACTIVE"] = "false"

%pip uninstall -q -y optimum optimum-intel
%pip install -q --extra-index-url https://download.pytorch.org/whl/cpu \
"llama-index" "llama-index-readers-file" "llama-index-llms-openvino>=0.2.2" "llama-index-embeddings-openvino>=0.2.0" "transformers>=4.40"

%pip install -q "git+https://github.com/huggingface/optimum-intel.git" \
"git+https://github.com/openvinotoolkit/nncf.git" \
"datasets" \
"accelerate"
%pip install --pre -Uq "openvino>=2024.2.0" openvino-tokenizers[transformers] --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly

WARNING: Skipping optimum as it is not installed.
WARNING: Skipping optimum-intel as it is not installed.
Note: you may need to restart the kernel to use updated packages.

from pathlib import Path
import requests
import io

text_example_en_path = Path("text_example_en.pdf")
text_example_en = "https://github.com/user-attachments/files/16171326/xeon6-e-cores-network-and-edge-brief.pdf"

if not text_example_en_path.exists():
    r = requests.get(url=text_example_en)
    content = io.BytesIO(r.content)
    with open("text_example_en.pdf", "wb") as f:
        f.write(content.read())

Download models

Download LLM

To run LLM locally, we have to download the model in the first step. It is possible to export your model to the OpenVINO IR format with the CLI, and load the model from local folder.

Large Language Models (LLMs) are a core component of agent. LlamaIndex does not serve its own LLMs, but rather provides a standard interface for interacting with many different LLMs. In this example, we can select Phi3-mini-instruct or Meta-Llama-3-8B-Instruct as LLM in agent pipeline. * phi3-mini-instruct - The Phi-3-Mini is a 3.8B parameters, lightweight, state-of-the-art open model trained with the Phi-3 datasets that includes both synthetic data and the filtered publicly available websites data with a focus on high-quality and reasoning dense properties. More details about model can be found in model card, Microsoft blog and technical report. * llama-3-8b-instruct - Llama 3 is an auto-regressive language model that uses an optimized transformer architecture. The tuned versions use supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to align with human preferences for helpfulness and safety. The Llama 3 instruction tuned models are optimized for dialogue use cases and outperform many of the available open source chat models on common industry benchmarks. More details about model can be found in Meta blog post, model website and model card. >Note: run model with demo, you will need to accept license agreement. >You must be a registered user in Hugging Face Hub. Please visit HuggingFace model card, carefully read terms of usage and click accept button. You will need to use an access token for the code below to run. For more information on access tokens, refer to this section of the documentation. >You can login on Hugging Face Hub in notebook environment, using following code:

## login to huggingfacehub to get access to pretrained model

from huggingface_hub import notebook_login, whoami

try:
    whoami()
    print('Authorization token already provided')
except OSError:
    notebook_login()

import ipywidgets as widgets

llm_model_ids = ["OpenVINO/Phi-3-mini-4k-instruct-int4-ov", "meta-llama/Meta-Llama-3-8B-Instruct"]

llm_model_id = widgets.Dropdown(
    options=llm_model_ids,
    value=llm_model_ids[0],
    description="Model:",
    disabled=False,
)

llm_model_id

Dropdown(description='Model:', options=('OpenVINO/Phi-3-mini-4k-instruct-int4-ov', 'meta-llama/Meta-Llama-3-8B…

from pathlib import Path
import huggingface_hub as hf_hub

llm_model_path = llm_model_id.value.split("/")[-1]
repo_name = llm_model_id.value.split("/")[0]

if not Path(llm_model_path).exists():
    if repo_name == "OpenVINO":
        hf_hub.snapshot_download(llm_model_id.value, local_dir=llm_model_path)
    else:
        !optimum-cli export openvino --model {llm_model_id.value} --task text-generation-with-past --trust-remote-code --weight-format int4 --group-size 128 --ratio 0.8 {llm_model_path}

Fetching 16 files:   0%|          | 0/16 [00:00<?, ?it/s]

.gitattributes:   0%|          | 0.00/1.52k [00:00<?, ?B/s]

config.json:   0%|          | 0.00/884 [00:00<?, ?B/s]

README.md:   0%|          | 0.00/3.75k [00:00<?, ?B/s]

openvino_detokenizer.xml:   0%|          | 0.00/3.25k [00:00<?, ?B/s]

generation_config.json:   0%|          | 0.00/172 [00:00<?, ?B/s]

configuration_phi3.py:   0%|          | 0.00/10.4k [00:00<?, ?B/s]

openvino_model.xml:   0%|          | 0.00/3.04M [00:00<?, ?B/s]

openvino_detokenizer.bin:   0%|          | 0.00/500k [00:00<?, ?B/s]

special_tokens_map.json:   0%|          | 0.00/569 [00:00<?, ?B/s]

added_tokens.json:   0%|          | 0.00/293 [00:00<?, ?B/s]

openvino_tokenizer.xml:   0%|          | 0.00/12.7k [00:00<?, ?B/s]

openvino_model.bin:   0%|          | 0.00/2.45G [00:00<?, ?B/s]

openvino_tokenizer.bin:   0%|          | 0.00/500k [00:00<?, ?B/s]

tokenizer.json:   0%|          | 0.00/1.84M [00:00<?, ?B/s]

tokenizer_config.json:   0%|          | 0.00/3.34k [00:00<?, ?B/s]

tokenizer.model:   0%|          | 0.00/500k [00:00<?, ?B/s]

Download Embedding model

Embedding model is another key component in RAG pipeline. It takes text as input, and return a long list of numbers used to capture the semantics of the text. An OpenVINO embedding model and tokenizer can be exported by feature-extraction task with optimum-cli. In this tutorial, we use bge-small-en-v1.5 as example.

embedding_model_id = "BAAI/bge-small-en-v1.5"
embedding_model_path = "bge-small-en-v1.5"

if not Path(embedding_model_path).exists():
    !optimum-cli export openvino --model {embedding_model_id} --task feature-extraction {embedding_model_path}

Create models

Create OpenVINO LLM

Select device for LLM model inference

import ipywidgets as widgets
import openvino as ov

core = ov.Core()

support_devices = core.available_devices

llm_device = widgets.Dropdown(
    options=support_devices + ["AUTO"],
    value="CPU",
    description="Device:",
    disabled=False,
)

llm_device

[ERROR] 09:10:57.134 [NPUBackends] Cannot find backend for inference. Make sure the device is available.

Dropdown(description='Device:', options=('CPU', 'AUTO'), value='CPU')

OpenVINO models can be run locally through the OpenVINOLLM class in LlamaIndex. If you have an Intel GPU, you can specify device_map="gpu" to run inference on it.

from llama_index.llms.openvino import OpenVINOLLM

ov_config = {"PERFORMANCE_HINT": "LATENCY", "NUM_STREAMS": "1", "CACHE_DIR": ""}


def completion_to_prompt(completion):
    return f"<|system|><|end|><|user|>{completion}<|end|><|assistant|>\n"


def messages_to_prompt(messages):
    prompt = ""
    for message in messages:
        if message.role == "system":
            prompt += f"<|system|>{message.content}<|end|>"
        elif message.role == "user":
            prompt += f"<|user|>{message.content}<|end|>"
        elif message.role == "assistant":
            prompt += f"<|assistant|>{message.content}<|end|>"

    # ensure we start with a system prompt, insert blank if needed
    if not prompt.startswith("<|system|>"):
        prompt = "<|system|><|end|>" + prompt

    # add final assistant prompt
    prompt = prompt + "<|assistant|>\n"

    return prompt


llm = OpenVINOLLM(
    model_id_or_path=str(llm_model_path),
    context_window=3900,
    max_new_tokens=1000,
    model_kwargs={"ov_config": ov_config},
    generate_kwargs={"do_sample": False, "temperature": None, "top_p": None},
    completion_to_prompt=completion_to_prompt,
    messages_to_prompt=messages_to_prompt,
    device_map=llm_device.value,
)

/home/ethan/intel/openvino_notebooks/openvino_env/lib/python3.11/site-packages/pydantic/_internal/_fields.py:161: UserWarning: Field "model_id" has conflict with protected namespace "model_".

You may be able to resolve this warning by setting model_config['protected_namespaces'] = ().
  warnings.warn(
Compiling the model to CPU ...

Create OpenVINO Embedding

Select device for embedding model inference

support_devices = core.available_devices

embedding_device = widgets.Dropdown(
    options=support_devices + ["AUTO"],
    value="CPU",
    description="Device:",
    disabled=False,
)

embedding_device

Dropdown(description='Device:', options=('CPU', 'AUTO'), value='CPU')

A Hugging Face embedding model can be supported by OpenVINO through OpenVINOEmbeddings class of LlamaIndex.

from llama_index.embeddings.huggingface_openvino import OpenVINOEmbedding

embedding = OpenVINOEmbedding(model_id_or_path=embedding_model_path, device=embedding_device.value)

Compiling the model to CPU ...

Create tools

In this examples, we will create 2 customized tools for multiply and add.

from llama_index.core.agent import ReActAgent
from llama_index.core.tools import FunctionTool


def multiply(a: float, b: float) -> float:
    """Multiply two numbers and returns the product"""
    return a * b


multiply_tool = FunctionTool.from_defaults(fn=multiply)


def divide(a: float, b: float) -> float:
    """Add two numbers and returns the sum"""
    return a / b


divide_tool = FunctionTool.from_defaults(fn=divide)

To demonstrate using RAG engines as a tool in an agent, we’re going to create a very simple RAG query engine as one of the tools.

Note: For a full RAG pipeline with OpenVINO, you can check the RAG notebooks

from llama_index.core import SimpleDirectoryReader
from llama_index.core import VectorStoreIndex, Settings

Settings.embed_model = embedding
Settings.llm = llm

reader = SimpleDirectoryReader(input_files=[text_example_en_path])
documents = reader.load_data()
index = VectorStoreIndex.from_documents(
    documents,
)

Now we turn our query engine into a tool by supplying the appropriate metadata (for the python functions, this was being automatically extracted so we didn’t need to add it):

from llama_index.core.tools import QueryEngineTool, ToolMetadata

vector_tool = QueryEngineTool(
    index.as_query_engine(streaming=True),
    metadata=ToolMetadata(
        name="vector_search",
        description="Useful for searching for basic facts about Intel Xeon 6 processors",
    ),
)

Run Agentic RAG

We modify our agent by adding this engine to our array of tools (we also remove the llm parameter, since it’s now provided by settings):

agent = ReActAgent.from_tools([multiply_tool, divide_tool, vector_tool], llm=llm, verbose=True)

Ask a question using multiple tools.

response = agent.chat("What's the maximum number of cores of 6731 sockets of Intel Xeon 6 processors ? Go step by step, using a tool to do any math.")

> Running step 49b4846c-74d1-4766-9321-06aa21df11e1. Step input: What's the maximum number of cores of 6731 sockets of Intel Xeon 6 processors ? Go step by step, using a tool to do any math.
Thought: The current language of the user is English. I need to use a tool to help me answer the question.
Action: vector_search
Action Input: {'input': 'maximum number of cores of Intel Xeon 6 processors'}
Observation: The Intel Xeon 6 processors with Efficient-cores have up to 144 cores per socket in 1- or 2-socket configurations.
> Running step cb9c570c-09b4-4ac6-9121-cd1ef67e63cc. Step input: None
Thought: I can answer without using any more tools. I'll use the user's language to answer.
Answer: The maximum number of cores for Intel Xeon 6 processors is 144 cores per socket.
Support: To calculate the maximum number of cores for 6731 sockets, you would multiply the number of cores per socket by the number of sockets.
Action: multiply
Action Input: {'a': 144, 'b': 6731}
Observation: 969264
> Running step 9601596a-63db-4791-92fe-750f3a0cd924. Step input: None
Thought: I can answer without using any more tools. I'll use the user's language to answer.
Answer: The maximum number of cores for 6731 sockets of Intel Xeon 6 processors is 969,264 cores.

agent.reset()