[Beta] Multi-modal models - LlamaIndex 🦙 v0.10.19

Concept#

Large language models (LLMs) are text-in, text-out. Large Multi-modal Models (LMMs) generalize this beyond the text modalities. For instance, models such as GPT-4V allow you to jointly input both images and text, and output text.

We’ve included a base MultiModalLLM abstraction to allow for text+image models. NOTE: This naming is subject to change!

Usage Pattern#

The following code snippet shows how you can get started using LMMs e.g. with GPT-4V.

from llama_index.multi_modal_llms.openai import OpenAIMultiModal
from llama_index.core.multi_modal_llms.generic_utils import load_image_urls
from llama_index.core import SimpleDirectoryReader

# load image documents from urls
image_documents = load_image_urls(image_urls)

# load image documents from local directory
image_documents = SimpleDirectoryReader(local_directory).load_data()

# non-streaming
openai_mm_llm = OpenAIMultiModal(
    model="gpt-4-vision-preview", api_key=OPENAI_API_TOKEN, max_new_tokens=300
)
response = openai_mm_llm.complete(
    prompt="what is in the image?", image_documents=image_documents
)

The following code snippet shows how you can build MultiModal Vector Stores/Index.

from llama_index.core.indices import MultiModalVectorStoreIndex
from llama_index.vector_stores.qdrant import QdrantVectorStore
from llama_index.core import SimpleDirectoryReader, StorageContext

import qdrant_client
from llama_index.core import SimpleDirectoryReader

# Create a local Qdrant vector store
client = qdrant_client.QdrantClient(path="qdrant_mm_db")

# if you only need image_store for image retrieval,
# you can remove text_sotre
text_store = QdrantVectorStore(
    client=client, collection_name="text_collection"
)
image_store = QdrantVectorStore(
    client=client, collection_name="image_collection"
)

storage_context = StorageContext.from_defaults(
    vector_store=text_store, image_store=image_store
)

# Load text and image documents from local folder
documents = SimpleDirectoryReader("./data_folder/").load_data()
# Create the MultiModal index
index = MultiModalVectorStoreIndex.from_documents(
    documents,
    storage_context=storage_context,
)

The following code snippet shows how you can use MultiModal Retriever and Query Engine.

from llama_index.multi_modal_llms.openai import OpenAIMultiModal
from llama_index.core import PromptTemplate
from llama_index.core.query_engine import SimpleMultiModalQueryEngine

retriever_engine = index.as_retriever(
    similarity_top_k=3, image_similarity_top_k=3
)

# retrieve more information from the GPT4V response
retrieval_results = retriever_engine.retrieve(response)

# if you only need image retrieval without text retrieval
# you can use `text_to_image_retrieve`
# retrieval_results = retriever_engine.text_to_image_retrieve(response)

qa_tmpl_str = (
    "Context information is below.\n"
    "---------------------\n"
    "{context_str}\n"
    "---------------------\n"
    "Given the context information and not prior knowledge, "
    "answer the query.\n"
    "Query: {query_str}\n"
    "Answer: "
)
qa_tmpl = PromptTemplate(qa_tmpl_str)

query_engine = index.as_query_engine(
    multi_modal_llm=openai_mm_llm, text_qa_template=qa_tmpl
)

query_str = "Tell me more about the Porsche"
response = query_engine.query(query_str)

Legend

✅ = should work fine
⚠️ = sometimes unreliable, may need more tuning to improve
🛑 = not available at the moment.

End to End Multi-Modal Work Flow#

The tables below attempt to show the initial steps with various LlamaIndex features for building your own Multi-Modal RAGs (Retrieval Augmented Generation). You can combine different modules/steps together for composing your own Multi-Modal RAG orchestration.

Query Type	Data Sources for MultiModal Vector Store/Index	MultiModal Embedding	Retriever	Query Engine	Output Data Type
Text ✅	Text ✅	Text ✅	Top-k retrieval ✅ Simple Fusion retrieval ✅	Simple Query Engine ✅	Retrieved Text ✅ Generated Text ✅
Image ✅	Image ✅	Image ✅ Image to Text Embedding ✅	Top-k retrieval ✅ Simple Fusion retrieval ✅	Simple Query Engine ✅	Retrieved Image ✅ Generated Image 🛑
Audio 🛑	Audio 🛑	Audio 🛑	🛑	🛑	Audio 🛑
Video 🛑	Video 🛑	Video 🛑	🛑	🛑	Video 🛑

Multi-Modal LLM Models#

These notebooks serve as examples how to leverage and integrate Multi-Modal LLM model, Multi-Modal embeddings, Multi-Modal vector stores, Retriever, Query engine for composing Multi-Modal Retrieval Augmented Generation (RAG) orchestration.

Multi-Modal Vision Models	Single Image Reasoning	Multiple Images Reasoning	Image Embeddings	Simple Query Engine	Pydantic Structured Output
GPT4V (OpenAI API)	✅	✅	🛑	✅	✅
GPT4V-Azure (Azure API)	✅	✅	🛑	✅	✅
Gemini (Google)	✅	✅	🛑	✅	✅
CLIP (Local host)	🛑	🛑	✅	🛑	🛑
LLaVa (replicate)	✅	🛑	🛑	✅	⚠️
Fuyu-8B (replicate)	✅	🛑	🛑	✅	⚠️
ImageBind [To integrate]	🛑	🛑	✅	🛑	🛑
MiniGPT-4	✅	🛑	🛑	✅	⚠️
CogVLM	✅	🛑	🛑	✅	⚠️
Qwen-VL [To integrate]	✅	🛑	🛑	✅	⚠️

Multi Modal Vector Stores#

Below table lists some vector stores supporting Multi-Modal use cases. Our LlamaIndex built-in MultiModalVectorStoreIndex supports building separate vector stores for image and text embedding vector stores. MultiModalRetriever, and SimpleMultiModalQueryEngine support text to text/image and image to image retrieval and simple ranking fusion functions for combining text and image retrieval results.

Multi-Modal Vector Stores	Single Vector Store	Multiple Vector Stores	Text Embedding	Image Embedding
LLamaIndex self-built MultiModal Index	🛑	✅	Can be arbitrary text embedding (Default is GPT3.5)	Can be arbitrary Image embedding (Default is CLIP)
Chroma	✅	🛑	CLIP ✅	CLIP ✅
Weaviate [To integrate]	✅	🛑	CLIP ✅ ImageBind ✅	CLIP ✅ ImageBind ✅

Multi-Modal LLM Modules#

We support integrations with GPT4-V, Anthropic (Opus, Sonnet), Gemini (Google), CLIP (OpenAI), BLIP (Salesforce), and Replicate (LLaVA, Fuyu-8B, MiniGPT-4, CogVLM), and more.

Multi-Modal Retrieval Augmented Generation#

We support Multi-Modal Retrieval Augmented Generation with different Multi-Modal LLMs with Multi-Modal vector stores.

Evaluation#

We support basic evaluation for Multi-Modal LLM and Retrieval Augmented Generation.

Evaluating Multi-Modal RAG

[Beta] Multi-modal models#