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.
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
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
%% 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
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 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 default compose.yaml
| MicroService | Open Source Project | HW | Port | Endpoint |
|---|---|---|---|---|
| Embedding | Langchain | Xeon | 6000 | /v1/embeddings |
| Retriever | Langchain, Redis | Xeon | 7000 | /v1/multimodal_retrieval |
| LVM | Langchain, TGI | Gaudi | 9399 | /v1/lvm |
| Dataprep | Redis, Langchain, TGI | Gaudi | 6007 | /v1/generate_transcripts, /v1/generate_captions, /v1/ingest_with_text |
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 |
| embedding | Gaudi | BridgeTower/bridgetower-large-itm-mlm-itc |
| LVM | Gaudi | llava-hf/llava-v1.6-vicuna-13b-hf |
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.
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.
To set up environment variables for deploying MultimodalQnA services, follow these steps:
-
Set the required environment variables:
# Example: export host_ip=$(hostname -I | awk '{print $1}') export host_ip="External_Public_IP" # Example: no_proxy="localhost, 127.0.0.1, 192.168.1.1" export no_proxy="Your_No_Proxy"
-
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"
-
Set up other environment variables:
Notice that you can only choose one command below to set up envs according to your hardware. Other that the port numbers may be set incorrectly.
# on Gaudi source ./docker_compose/intel/hpu/gaudi/set_env.sh # on Xeon source ./docker_compose/intel/cpu/xeon/set_env.sh
Refer to the Gaudi Guide to build docker images from source.
Find the corresponding compose.yaml.
cd GenAIExamples/MultimodalQnA/docker_compose/intel/hpu/gaudi/
docker compose -f compose.yaml up -dNotice: Currently only the Habana Driver 1.17.x is supported for Gaudi.
Refer to the Xeon Guide for more instructions on building docker images from source.
Find the corresponding compose.yaml.
cd GenAIExamples/MultimodalQnA/docker_compose/intel/cpu/xeon/
docker compose -f compose.yaml up -d
