GenAIExamples/MultimodalQnA

MultimodalQnA Application

Suppose you possess a set of videos, images, audio files, PDFs, or some combination thereof and wish to perform question-answering to extract insights from these documents. To respond to your questions, the system needs to comprehend a mix of textual, visual, and audio facts drawn from the document contents. The MultimodalQnA framework offers an optimal solution for this purpose.

MultimodalQnA addresses your questions by dynamically fetching the most pertinent multimodal information (e.g. images, transcripts, and captions) from your collection of video, image, audio, and PDF files. For this purpose, MultimodalQnA utilizes BridgeTower model, a multimodal encoding transformer model which merges visual and textual data into a unified semantic space. During the ingestion phase, the BridgeTower model embeds both visual cues and auditory facts as texts, and those embeddings are then stored in a vector database. When it comes to answering a question, the MultimodalQnA will fetch its most relevant multimodal content from the vector store and feed it into a downstream Large Vision-Language Model (LVM) as input context to generate a response for the user, which can be text or audio.

The MultimodalQnA architecture shows below:

MultimodalQnA is implemented on top of GenAIComps, the MultimodalQnA Flow Chart shows below:

---
config:
  flowchart:
    nodeSpacing: 400
    rankSpacing: 100
    curve: linear
  themeVariables:
    fontSize: 50px
---
flowchart LR
    %% Colors %%
    classDef blue fill:#ADD8E6,stroke:#ADD8E6,stroke-width:2px,fill-opacity:0.5
    classDef orange fill:#FBAA60,stroke:#ADD8E6,stroke-width:2px,fill-opacity:0.5
    classDef orchid fill:#C26DBC,stroke:#ADD8E6,stroke-width:2px,fill-opacity:0.5
    classDef invisible fill:transparent,stroke:transparent;
    style MultimodalQnA-MegaService stroke:#000000
    %% Subgraphs %%
    subgraph MultimodalQnA-MegaService["MultimodalQnA-MegaService"]
        direction LR
        EM([Embedding <br>]):::blue
        RET([Retrieval <br>]):::blue
        LVM([LVM <br>]):::blue
    end
    subgraph UserInterface[" User Interface "]
        direction LR
        a([User Input Query]):::orchid
        Ingest([Ingest data]):::orchid
        UI([UI server<br>]):::orchid
    end

    ASR{{Whisper service <br>}}
    TEI_EM{{Embedding service <br>}}
    VDB{{Vector DB<br><br>}}
    R_RET{{Retriever service <br>}}
    DP([Data Preparation<br>]):::blue
    LVM_gen{{LVM Service <br>}}
    GW([MultimodalQnA GateWay<br>]):::orange
    TTS{{SpeechT5 service <br>}}

    %% Data Preparation flow
    %% Ingest data flow
    direction LR
    Ingest[Ingest data] --> UI
    UI -->DP
    DP <-.-> TEI_EM

    %% Questions interaction
    direction LR
    a[User Input Query] --> UI
    UI --> GW
    GW <==> MultimodalQnA-MegaService
    EM ==> RET
    RET ==> LVM

    %% Embedding service flow
    direction LR
    EM <-.-> TEI_EM
    RET <-.-> R_RET
    LVM <-.-> LVM_gen

    direction TB
    %% Vector DB interaction
    R_RET <-.->VDB
    DP <-.->VDB

    %% Audio speech recognition used for translating audio queries to text
    GW <-.-> ASR

    %% Generate spoken responses with text-to-speech using the SpeechT5 model
    GW <-.-> TTS

This MultimodalQnA use case performs Multimodal-RAG using LangChain, Redis VectorDB and Text Generation Inference on Intel Gaudi2 and Intel Xeon Scalable Processors, and we invite contributions from other hardware vendors to expand the example.

The Whisper Service is used by MultimodalQnA for converting audio queries to text. If a spoken response is requested, the SpeechT5 Service translates the text response from the LVM to a speech audio file.

The Intel Gaudi2 accelerator supports both training and inference for deep learning models in particular for LLMs. Visit Habana AI products for more details.

In the below, we provide a table that describes for each microservice component in the MultimodalQnA architecture, the default configuration of the open source project, hardware, port, and endpoint.

Gaudi and Xeon default compose.yaml settings

MicroService	Open Source Project	HW	Port	Endpoint
Dataprep	Redis, Langchain, TGI	Xeon	6007	/v1/generate_transcripts, /v1/generate_captions, /v1/ingest
Embedding	Langchain	Xeon	6000	/v1/embeddings
LVM	Langchain, Transformers	Xeon	9399	/v1/lvm
Retriever	Langchain, Redis	Xeon	7000	/v1/retrieval
SpeechT5	Transformers	Xeon	7055	/v1/tts
Whisper	Transformers	Xeon	7066	/v1/asr
Dataprep	Redis, Langchain, TGI	Gaudi	6007	/v1/generate_transcripts, /v1/generate_captions, /v1/ingest
Embedding	Langchain	Gaudi	6000	/v1/embeddings
LVM	Langchain, TGI	Gaudi	9399	/v1/lvm
Retriever	Langchain, Redis	Gaudi	7000	/v1/retrieval
SpeechT5	Transformers	Gaudi	7055	/v1/tts
Whisper	Transformers	Gaudi	7066	/v1/asr

Required Models

By default, the embedding and LVM models are set to a default value as listed below:

Service	HW	Model
embedding	Xeon	BridgeTower/bridgetower-large-itm-mlm-itc
LVM	Xeon	llava-hf/llava-1.5-7b-hf
SpeechT5	Xeon	microsoft/speecht5_tts
Whisper	Xeon	openai/whisper-small
embedding	Gaudi	BridgeTower/bridgetower-large-itm-mlm-itc
LVM	Gaudi	llava-hf/llava-v1.6-vicuna-13b-hf
SpeechT5	Gaudi	microsoft/speecht5_tts
Whisper	Gaudi	openai/whisper-small

You can choose other LVM models, such as llava-hf/llava-1.5-7b-hf and llava-hf/llava-1.5-13b-hf, as needed.

Deploy MultimodalQnA Service

The MultimodalQnA service can be effortlessly deployed on either Intel Gaudi2 or Intel XEON Scalable Processors.

Currently we support deploying MultimodalQnA services with docker compose. The docker_compose directory has folders which include compose.yaml files for different hardware types:

📂 docker_compose
├── 📂 amd
│   └── 📂 gpu
│       └── 📂 rocm
│           ├── 📄 compose.yaml
│           └── ...
└── 📂 intel
    ├── 📂 cpu
    │   └── 📂 xeon
    │       ├── 📄 compose.yaml
    │       └── ...
    └── 📂 hpu
        └── 📂 gaudi
            ├── 📄 compose.yaml
            └── ...

Setup Environment Variables

To set up environment variables for deploying MultimodalQnA services, follow these steps:

1. Set the required environment variables:

   # Example: export host_ip=$(hostname -I | awk '{print $1}')
   export host_ip="External_Public_IP"
   
   # Append the host_ip to the no_proxy list to allow container communication
   # Example: no_proxy="localhost, 127.0.0.1, 192.168.1.1"
   export no_proxy="${no_proxy},${host_ip}"
   

2. If you are in a proxy environment, also set the proxy-related environment variables:

   export http_proxy="Your_HTTP_Proxy"
   export https_proxy="Your_HTTPs_Proxy"
   

3. Set up other environment variables:

   Choose one command below to set env vars according to your hardware. Otherwise, the port numbers may be set incorrectly.

   # on Gaudi
   cd docker_compose/intel/hpu/gaudi
   source ./set_env.sh
   
   # on Xeon
   cd docker_compose/intel/cpu/xeon
   source ./set_env.sh
   

Deploy MultimodalQnA on Gaudi

Refer to the Gaudi Guide if you would like to build docker images from source, otherwise images will be pulled from Docker Hub.

Find the corresponding compose.yaml.

# While still in the docker_compose/intel/hpu/gaudi directory, use docker compose to bring up the services
docker compose -f compose.yaml up -d

Notice: Currently only the Habana Driver 1.18.x is supported for Gaudi.

Deploy MultimodalQnA on Xeon

Refer to the Xeon Guide if you would like to build docker images from source, otherwise images will be pulled from Docker Hub.

Find the corresponding compose.yaml.

# While still in the docker_compose/intel/cpu/xeon directory, use docker compose to bring up the services
docker compose -f compose.yaml up -d

MultimodalQnA Demo on Gaudi2

Multimodal QnA UI

Video Ingestion

Text Query following the ingestion of a Video

Image Ingestion

Text Query following the ingestion of an image

Text Query following the ingestion of an image using text-to-speech

Audio Ingestion

Text Query following the ingestion of an Audio Podcast

PDF Ingestion

Text query following the ingestion of a PDF

View, Refresh, and Delete ingested media in the Vector Store