Create a RAG system using OpenVINO and LlamaIndex#

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

Retrieval-augmented generation (RAG) is a technique for augmenting LLM knowledge with additional, often private or real-time, data. LLMs can reason about wide-ranging topics, but their knowledge is limited to the public data up to a specific point in time that they were trained on. If you want to build AI applications that can reason about private data or data introduced after a model’s cutoff date, you need to augment the knowledge of the model with the specific information it needs. The process of bringing the appropriate information and inserting it into the model prompt is known as Retrieval Augmented Generation (RAG).

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.

The tutorial consists of the following steps:

Install prerequisites
Download and convert the model from a public source using the OpenVINO integration with Hugging Face Optimum.
Compress model weights to 4-bit or 8-bit data types using NNCF
Create a RAG chain pipeline
Run Q&A pipeline

In this example, the customized RAG pipeline consists of following components in order, where embedding, rerank and LLM will be deployed with OpenVINO to optimize their inference performance.

Table of contents:

Prerequisites
Select model for inference
login to huggingfacehub to get access to pretrained model
Convert model and compress model weights
Select device for inference and model variant
Load model
Run QA over Document

Prerequisites#

Install required dependencies

import os

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

%pip install -Uq pip
%pip uninstall -q -y optimum optimum-intel
%pip install -q --extra-index-url https://download.pytorch.org/whl/cpu\
"llama-index" "pymupdf" "llama-index-readers-file" "llama-index-vector-stores-faiss" "llama-index-llms-openvino" "llama-index-embeddings-openvino" "llama-index-postprocessor-openvino-rerank" "transformers>=4.40"\
"git+https://github.com/huggingface/optimum-intel.git"\
"git+https://github.com/openvinotoolkit/nncf.git"\
"datasets"\
"accelerate"\
"gradio"
%pip install --pre -Uq openvino openvino-tokenizers[transformers] --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly

Note: you may need to restart the kernel to use updated packages.
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.
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
langchain-huggingface 0.0.1 requires huggingface-hub>=0.23.0, but you have huggingface-hub 0.20.3 which is incompatible.
Note: you may need to restart the kernel to use updated packages.
Note: you may need to restart the kernel to use updated packages.

import os
from pathlib import Path
import requests
import shutil
import io

# fetch model configuration

config_shared_path = Path("../../utils/llm_config.py")
config_dst_path = Path("llm_config.py")
text_example_en_path = Path("text_example_en.pdf")
text_example_cn_path = Path("text_example_cn.pdf")
text_example_en = "https://github.com/openvinotoolkit/openvino_notebooks/files/15039728/Platform.Brief_Intel.vPro.with.Intel.Core.Ultra_Final.pdf"
text_example_cn = "https://github.com/openvinotoolkit/openvino_notebooks/files/15039713/Platform.Brief_Intel.vPro.with.Intel.Core.Ultra_Final_CH.pdf"

if not config_dst_path.exists():
    if config_shared_path.exists():
        try:
            os.symlink(config_shared_path, config_dst_path)
        except Exception:
            shutil.copy(config_shared_path, config_dst_path)
    else:
        r = requests.get(url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/llm_config.py")
        with open("llm_config.py", "w", encoding="utf-8") as f:
            f.write(r.text)
elif not os.path.islink(config_dst_path):
    print("LLM config will be updated")
    if config_shared_path.exists():
        shutil.copy(config_shared_path, config_dst_path)
    else:
        r = requests.get(url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/llm_config.py")
        with open("llm_config.py", "w", encoding="utf-8") as f:
            f.write(r.text)


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())

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

LLM config will be updated

Select model for inference#

The tutorial supports different models, you can select one from the provided options to compare the quality of open source LLM solutions.

Note: conversion of some models can require additional actions from user side and at least 64GB RAM for conversion.

The available embedding model options are:

BGE embedding is a general Embedding Model. The model is pre-trained using RetroMAE and trained on large-scale pair data using contrastive learning.

The available rerank model options are:

Reranker model with cross-encoder will perform full-attention over the input pair, which is more accurate than embedding model (i.e., bi-encoder) but more time-consuming than embedding model. Therefore, it can be used to re-rank the top-k documents returned by embedding model.

You can also find available LLM model options in llm-chatbot notebook.

from pathlib import Path
import openvino as ov
import ipywidgets as widgets

Convert model and compress model weights#

The Weights Compression algorithm is aimed at compressing the weights of the models and can be used to optimize the model footprint and performance of large models where the size of weights is relatively larger than the size of activations, for example, Large Language Models (LLM). Compared to INT8 compression, INT4 compression improves performance even more, but introduces a minor drop in prediction quality.

from llm_config import (
    SUPPORTED_EMBEDDING_MODELS,
    SUPPORTED_RERANK_MODELS,
    SUPPORTED_LLM_MODELS,
)

model_languages = list(SUPPORTED_LLM_MODELS)

model_language = widgets.Dropdown(
    options=model_languages,
    value=model_languages[0],
    description="Model Language:",
    disabled=False,
)

model_language

Dropdown(description='Model Language:', options=('English', 'Chinese', 'Japanese'), value='English')

llm_model_ids = [model_id for model_id, model_config in SUPPORTED_LLM_MODELS[model_language.value].items() if model_config.get("rag_prompt_template")]

llm_model_id = widgets.Dropdown(
    options=llm_model_ids,
    value=llm_model_ids[-1],
    description="Model:",
    disabled=False,
)

llm_model_id

Dropdown(description='Model:', index=12, options=('tiny-llama-1b-chat', 'gemma-2b-it', 'red-pajama-3b-chat', '…

llm_model_configuration = SUPPORTED_LLM_MODELS[model_language.value][llm_model_id.value]
print(f"Selected LLM model {llm_model_id.value}")

Selected LLM model neural-chat-7b-v3-1

Optimum Intel is the interface between the Transformers and Diffusers libraries and OpenVINO to accelerate end-to-end pipelines on Intel architectures. It provides ease-to-use cli interface for exporting models to OpenVINO Intermediate Representation (IR) format.

The command bellow demonstrates basic command for model export with optimum-cli

optimum-cli export openvino --model <model_id_or_path> --task <task> <out_dir>

where --model argument is model id from HuggingFace Hub or local directory with model (saved using .save_pretrained method), --task is one of supported task that exported model should solve. For LLMs it will be text-generation-with-past. If model initialization requires to use remote code, --trust-remote-code flag additionally should be passed.

LLM conversion and Weights Compression using Optimum-CLI#

You can also apply fp16, 8-bit or 4-bit weight compression on the Linear, Convolutional and Embedding layers when exporting your model with the CLI by setting --weight-format to respectively fp16, int8 or int4. This type of optimization allows to reduce the memory footprint and inference latency. By default the quantization scheme for int8/int4 will be asymmetric, to make it symmetric you can add --sym.

For INT4 quantization you can also specify the following arguments :

The --group-size parameter will define the group size to use for quantization, -1 it will results in per-column quantization.
The --ratio parameter controls the ratio between 4-bit and 8-bit quantization. If set to 0.9, it means that 90% of the layers will be quantized to int4 while 10% will be quantized to int8.

Smaller group_size and ratio values usually improve accuracy at the sacrifice of the model size and inference latency.

Note: There may be no speedup for INT4/INT8 compressed models on dGPU.

from IPython.display import Markdown, display

prepare_int4_model = widgets.Checkbox(
    value=True,
    description="Prepare INT4 model",
    disabled=False,
)
prepare_int8_model = widgets.Checkbox(
    value=False,
    description="Prepare INT8 model",
    disabled=False,
)
prepare_fp16_model = widgets.Checkbox(
    value=False,
    description="Prepare FP16 model",
    disabled=False,
)

display(prepare_int4_model)
display(prepare_int8_model)
display(prepare_fp16_model)

Checkbox(value=True, description='Prepare INT4 model')

Checkbox(value=False, description='Prepare INT8 model')

Checkbox(value=False, description='Prepare FP16 model')

Weight compression with AWQ#

Activation-aware Weight Quantization (AWQ) is an algorithm that tunes model weights for more accurate INT4 compression. It slightly improves generation quality of compressed LLMs, but requires significant additional time for tuning weights on a calibration dataset. We use wikitext-2-raw-v1/train subset of the Wikitext dataset for calibration.

Below you can enable AWQ to be additionally applied during model export with INT4 precision.

Note: Applying AWQ requires significant memory and time.

Note: It is possible that there will be no matching patterns in the model to apply AWQ, in such case it will be skipped.

enable_awq = widgets.Checkbox(
    value=False,
    description="Enable AWQ",
    disabled=not prepare_int4_model.value,
)
display(enable_awq)

Checkbox(value=False, description='Enable AWQ')

pt_model_id = llm_model_configuration["model_id"]
pt_model_name = llm_model_id.value.split("-")[0]
fp16_model_dir = Path(llm_model_id.value) / "FP16"
int8_model_dir = Path(llm_model_id.value) / "INT8_compressed_weights"
int4_model_dir = Path(llm_model_id.value) / "INT4_compressed_weights"


def convert_to_fp16():
    if (fp16_model_dir / "openvino_model.xml").exists():
        return
    remote_code = llm_model_configuration.get("remote_code", False)
    export_command_base = "optimum-cli export openvino --model {} --task text-generation-with-past --weight-format fp16".format(pt_model_id)
    if remote_code:
        export_command_base += " --trust-remote-code"
    export_command = export_command_base + " " + str(fp16_model_dir)
    display(Markdown("**Export command:**"))
    display(Markdown(f"`{export_command}`"))
    ! $export_command


def convert_to_int8():
    if (int8_model_dir / "openvino_model.xml").exists():
        return
    int8_model_dir.mkdir(parents=True, exist_ok=True)
    remote_code = llm_model_configuration.get("remote_code", False)
    export_command_base = "optimum-cli export openvino --model {} --task text-generation-with-past --weight-format int8".format(pt_model_id)
    if remote_code:
        export_command_base += " --trust-remote-code"
    export_command = export_command_base + " " + str(int8_model_dir)
    display(Markdown("**Export command:**"))
    display(Markdown(f"`{export_command}`"))
    ! $export_command


def convert_to_int4():
    compression_configs = {
        "zephyr-7b-beta": {
            "sym": True,
            "group_size": 64,
            "ratio": 0.6,
        },
        "mistral-7b": {
            "sym": True,
            "group_size": 64,
            "ratio": 0.6,
        },
        "minicpm-2b-dpo": {
            "sym": True,
            "group_size": 64,
            "ratio": 0.6,
        },
        "gemma-2b-it": {
            "sym": True,
            "group_size": 64,
            "ratio": 0.6,
        },
        "notus-7b-v1": {
            "sym": True,
            "group_size": 64,
            "ratio": 0.6,
        },
        "neural-chat-7b-v3-1": {
            "sym": True,
            "group_size": 64,
            "ratio": 0.6,
        },
        "llama-2-chat-7b": {
            "sym": True,
            "group_size": 128,
            "ratio": 0.8,
        },
        "llama-3-8b-instruct": {
            "sym": True,
            "group_size": 128,
            "ratio": 0.8,
        },
        "gemma-7b-it": {
            "sym": True,
            "group_size": 128,
            "ratio": 0.8,
        },
        "chatglm2-6b": {
            "sym": True,
            "group_size": 128,
            "ratio": 0.72,
        },
        "qwen-7b-chat": {"sym": True, "group_size": 128, "ratio": 0.6},
        "red-pajama-3b-chat": {
            "sym": False,
            "group_size": 128,
            "ratio": 0.5,
        },
        "default": {
            "sym": False,
            "group_size": 128,
            "ratio": 0.8,
        },
    }

    model_compression_params = compression_configs.get(llm_model_id.value, compression_configs["default"])
    if (int4_model_dir / "openvino_model.xml").exists():
        return
    remote_code = llm_model_configuration.get("remote_code", False)
    export_command_base = "optimum-cli export openvino --model {} --task text-generation-with-past --weight-format int4".format(pt_model_id)
    int4_compression_args = " --group-size {} --ratio {}".format(model_compression_params["group_size"], model_compression_params["ratio"])
    if model_compression_params["sym"]:
        int4_compression_args += " --sym"
    if enable_awq.value:
        int4_compression_args += " --awq --dataset wikitext2 --num-samples 128"
    export_command_base += int4_compression_args
    if remote_code:
        export_command_base += " --trust-remote-code"
    export_command = export_command_base + " " + str(int4_model_dir)
    display(Markdown("**Export command:**"))
    display(Markdown(f"`{export_command}`"))
    ! $export_command


if prepare_fp16_model.value:
    convert_to_fp16()
if prepare_int8_model.value:
    convert_to_int8()
if prepare_int4_model.value:
    convert_to_int4()

Let’s compare model size for different compression types

fp16_weights = fp16_model_dir / "openvino_model.bin"
int8_weights = int8_model_dir / "openvino_model.bin"
int4_weights = int4_model_dir / "openvino_model.bin"

if fp16_weights.exists():
    print(f"Size of FP16 model is {fp16_weights.stat().st_size / 1024 / 1024:.2f} MB")
for precision, compressed_weights in zip([8, 4], [int8_weights, int4_weights]):
    if compressed_weights.exists():
        print(f"Size of model with INT{precision} compressed weights is {compressed_weights.stat().st_size / 1024 / 1024:.2f} MB")
    if compressed_weights.exists() and fp16_weights.exists():
        print(f"Compression rate for INT{precision} model: {fp16_weights.stat().st_size / compressed_weights.stat().st_size:.3f}")

Size of model with INT4 compressed weights is 5069.90 MB

Convert embedding model using Optimum-CLI#

Since some embedding models can only support limited languages, we can filter them out according the LLM you selected.

embedding_model_id = list(SUPPORTED_EMBEDDING_MODELS[model_language.value])

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

embedding_model_id

Dropdown(description='Embedding Model:', options=('bge-small-en-v1.5', 'bge-large-en-v1.5'), value='bge-small-…

embedding_model_configuration = SUPPORTED_EMBEDDING_MODELS[model_language.value][embedding_model_id.value]
print(f"Selected {embedding_model_id.value} model")

Selected bge-small-en-v1.5 model

OpenVINO embedding model and tokenizer can be exported by feature-extraction task with optimum-cli.

export_command_base = "optimum-cli export openvino --model {} --task feature-extraction".format(embedding_model_configuration["model_id"])
export_command = export_command_base + " " + str(embedding_model_id.value)

if not Path(embedding_model_id.value).exists():
    ! $export_command

Convert rerank model using Optimum-CLI#

rerank_model_id = list(SUPPORTED_RERANK_MODELS)

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

rerank_model_id

Dropdown(description='Rerank Model:', options=('bge-reranker-large', 'bge-reranker-base'), value='bge-reranker…

rerank_model_configuration = SUPPORTED_RERANK_MODELS[rerank_model_id.value]
print(f"Selected {rerank_model_id.value} model")

Selected bge-reranker-large model

Since rerank model is sort of sentence classification task, its OpenVINO IR and tokenizer can be exported by text-classification task with optimum-cli.

export_command_base = "optimum-cli export openvino --model {} --task text-classification".format(rerank_model_configuration["model_id"])
export_command = export_command_base + " " + str(rerank_model_id.value)

if not Path(rerank_model_id.value).exists():
    ! $export_command

Select device for inference and model variant#

Note: There may be no speedup for INT4/INT8 compressed models on dGPU.

Select device for embedding model inference#

core = ov.Core()

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', 'GPU', 'AUTO'), value='CPU')

print(f"Embedding model will be loaded to {embedding_device.value} device for text embedding")

Embedding model will be loaded to CPU device for text embedding

Select device for rerank model inference#

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

rerank_device

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

print(f"Rerenk model will be loaded to {rerank_device.value} device for text reranking")

Rerenk model will be loaded to CPU device for text reranking

Select device for LLM model inference#

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

llm_device

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

print(f"LLM model will be loaded to {llm_device.value} device for response generation")

LLM model will be loaded to CPU device for response generation

Load models#

Load embedding model#

Now 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(folder_name=embedding_model_id.value, device=embedding_device.value)

embeddings = embedding.get_text_embedding("Hello World!")
print(len(embeddings))
print(embeddings[:5])

Compiling the model to CPU ...

384
[-0.003275666618719697, -0.01169075071811676, 0.04155930131673813, -0.03814813867211342, 0.02418304793536663]

Load rerank model#

Now a Hugging Face embedding model can be supported by OpenVINO through OpenVINORerank class of LlamaIndex.

Note: Rerank can be skipped in RAG.

from llama_index.postprocessor.openvino_rerank import OpenVINORerank

reranker = OpenVINORerank(model=rerank_model_id.value, device=rerank_device.value, top_n=2)

Compiling the model to CPU ...

Load LLM model#

OpenVINO models can be run locally through the HuggingFacePipeline class. To deploy a model with OpenVINO, you can specify the backend="openvino" parameter to trigger OpenVINO as backend inference framework.

available_models = []
if int4_model_dir.exists():
    available_models.append("INT4")
if int8_model_dir.exists():
    available_models.append("INT8")
if fp16_model_dir.exists():
    available_models.append("FP16")

model_to_run = widgets.Dropdown(
    options=available_models,
    value=available_models[0],
    description="Model to run:",
    disabled=False,
)

model_to_run

Dropdown(description='Model to run:', options=('INT4',), value='INT4')

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

if model_to_run.value == "INT4":
    model_dir = int4_model_dir
elif model_to_run.value == "INT8":
    model_dir = int8_model_dir
else:
    model_dir = fp16_model_dir
print(f"Loading model from {model_dir}")

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

if "GPU" in llm_device.value and "qwen2-7b-instruct" in llm_model_id.value:
    ov_config["GPU_ENABLE_SDPA_OPTIMIZATION"] = "NO"

# On a GPU device a model is executed in FP16 precision. For red-pajama-3b-chat model there known accuracy
# issues caused by this, which we avoid by setting precision hint to "f32".
if llm_model_id.value == "red-pajama-3b-chat" and "GPU" in core.available_devices and llm_device.value in ["GPU", "AUTO"]:
    ov_config["INFERENCE_PRECISION_HINT"] = "f32"

llm = OpenVINOLLM(
    model_name=str(model_dir),
    tokenizer_name=str(model_dir),
    context_window=3900,
    max_new_tokens=2,
    model_kwargs={"ov_config": ov_config, "trust_remote_code": True},
    generate_kwargs={"temperature": 0.7, "top_k": 50, "top_p": 0.95},
    device_map=llm_device.value,
)

response = llm.complete("2 + 2 =")
print(str(response))

The argument trust_remote_code is to be used along with export=True. It will be ignored.

Loading model from neural-chat-7b-v3-1/INT4_compressed_weights

Compiling the model to CPU ...
/home/ethan/intel/openvino_notebooks/openvino_env/lib/python3.10/site-packages/transformers/generation/configuration_utils.py:492: UserWarning: do_sample is set to False. However, temperature is set to 0.7 -- this flag is only used in sample-based generation modes. You should set do_sample=True or unset temperature.
  warnings.warn(
/home/ethan/intel/openvino_notebooks/openvino_env/lib/python3.10/site-packages/transformers/generation/configuration_utils.py:497: UserWarning: do_sample is set to False. However, top_p is set to 0.95 -- this flag is only used in sample-based generation modes. You should set do_sample=True or unset top_p.
  warnings.warn(
Setting pad_token_id to eos_token_id:2 for open-end generation.

Run QA over Document#

A typical RAG application has two main components:

Indexing: a pipeline for ingesting data from a source and indexing it. This usually happen offline.
Retrieval and generation: the actual RAG chain, which takes the user query at run time and retrieves the relevant data from the index, then passes that to the model.

The most common full sequence from raw data to answer looks like:

Indexing

Load: First we need to load our data. We’ll use DocumentLoaders for this.
Split: Text splitters break large Documents into smaller chunks. This is useful both for indexing data and for passing it in to a model, since large chunks are harder to search over and won’t in a model’s finite context window.
Store: We need somewhere to store and index our splits, so that they can later be searched over. This is often done using a VectorStore and Embeddings model.

from llama_index.core import VectorStoreIndex, StorageContext
from llama_index.core.node_parser import SentenceSplitter
from llama_index.core import Settings
from llama_index.readers.file import PyMuPDFReader
from llama_index.vector_stores.faiss import FaissVectorStore
import faiss

if model_language.value == "English":
    text_example_path = "text_example_en.pdf"
else:
    text_example_path = "text_example_cn.pdf"

loader = PyMuPDFReader()
documents = loader.load(file_path=text_example_path)

# dimensions of embedding model
d = embedding._model.request.outputs[0].get_partial_shape()[2].get_length()
faiss_index = faiss.IndexFlatL2(d)
Settings.embed_model = embedding
llm.max_new_tokens = 2048
Settings.llm = llm


vector_store = FaissVectorStore(faiss_index=faiss_index)

Retrieval and generation

Retrieve: Given a user input, relevant splits are retrieved from storage using a Retriever.
Generate: A LLM produces an answer using a prompt that includes the question and the retrieved data.

storage_context = StorageContext.from_defaults(vector_store=vector_store)
index = VectorStoreIndex.from_documents(
    documents,
    storage_context=storage_context,
    transformations=[SentenceSplitter(chunk_size=500, chunk_overlap=50)],
)
query_engine = index.as_query_engine(streaming=True, similarity_top_k=10, node_postprocessors=[reranker])
if model_language.value == "English":
    query = "What can Intel vPro® Enterprise systems offer?"
else:
    query = "英特尔博锐® Enterprise系统提供哪些功能？"

streaming_response = query_engine.query(query)
streaming_response.print_response_stream()

Setting pad_token_id to eos_token_id:2 for open-end generation.

Intel vPro® Enterprise systems offer a range of features and capabilities, including dynamic root of trust, system management mode (SMM) protections, memory encryption with multi-key support, OS kernel protection, out-of-band management with remote KVM control, unique device identifier, device history, and in-band manageability plug-ins. These features are part of the strong portfolio of security and manageability technologies that form the foundation of the Intel vPro platform, which delivers differentiated capabilities to organizations of all sizes.