An exemplary project to demonstrate the art-of-the-possible for informative video search using off-the-shelf ML models, Elasticsearch ELSER semantic search, Elasticsearch hosted ML models, Elasticsearch non-text-media kNN search, and Elasticsearch Reciprocal Rank Fusion (RRF)!
Modern organizations, from large enterprises to school districts, share a lot of pertinent and timely information through video content (e.g., Town Halls, Tutorials, Webinars, Board Meetings). Typically such content addresses a wide range of topics within a given video, with each topic of interest only to a specific subset of the audience. Without a good intra-video search option, viewers are forced to consume such content linearly. At best, this potentially wastes a viewer's valuable time; at worst, it creates enough of a barrier to consumption that pertinent information (which could be a catalyst to further productivity) goes unwatched.
In general, few options exist to make informative, often private (e.g., internal use only), video practically searchable. Some existing projects summarize video content using a public Speech-To-Text (STT) engine and LLM (e.g., as offered by OpenAI), and then offer search on top of the summarized output. Use of such services, however, potentially puts private information at risk for public exposure. Further, and arguably more importantly, speakers in informative video are generally:
- subject matter experts
- well-spoken
- already producing expertly-curated summarized content
Adding a LLM to summarize their carefully chosen words can easily skew their intended meaning. As an example, consider a company townhall meeting which includes a discussion from the VP of HR regarding a layoff: the words spoken have been carefully selected and their nuance matters. This project (as the name implies) intentionally forgoes such paraphrasing and instead lets subject matter experts speak for themselves.
Ingest functions run once per title ingested on a compute node external to Elasticsearch.
- Convert audio to paragraphical text
- Extract audio from file
- Generate text sentences from speech (STT)
- Assign speakers to sentences (diarization)
- Generate embedding for voice
- Search for embedding in Elasticsearch using kNN
- If no match, create new voice record in Elasticsearch
- Group sentences
- By speaker
- By "thought" (using sentence similarity model)
- Split if greater than ELSER token limit
- Convert slides (if present) to keywords
- Extract video frames from file
- Determine unique frames through image similarity comparison
- Apply Optical Character Recognition (OCR) to unique frames
- Keep only unique, high-confidence frame text
- Align and merge extracted frame text with "thoughts" (from step 1.4) to create clauses
- Store clauses in Elasticsearch
- Apply ELSER encoding to clause text
Search functions are driven by a small external compute instance running Streamlit. Search operations are conducted on Elasticsearch.
- Extract keywords from query
- Submit query (ELSER against clause text) and keywords (BM25 against clause text and associated frame text) to Elasticsearch for a hybrid query (this is used to obtain a minimum score; if it is below a threshold, return no results)
- Submit query (ELSER against clause text) and keywords (BM25 against clause text and associated frame text) to Elasticsearch for a RRF query
- Run RRF results through a Q&A model (running on Elasticsearch ML nodes) to highlight answer
Some of the models require a Hugging Face token and acceptance of respective user agreements.
- Create a Hugging Face account
- Generate a token
- Accept the user agreement for the following models: Segmentation, Voice Activity Detection (VAD), and Speaker Diarization.
You will need a modern Elastic (>= 8.9) instance with sufficient ML resources (at least 4GB of RAM) to support this demo. You can start a 2 week free trial of Elastic Cloud here. Be sure to select "4 GB RAM" for the Machine Learning instance type. When you create the instance, you will be provided a password for the elastic user; make note of it. Additionally, from the Deployments page, you will need the endpoint of your Elasticsearch instance.
Because many content sites do not permit anonymous, external playback of video, this demo will upload your ingested video to (s3, gcs) cloud bucket storage to enable the UI to later play it back. You will need to setup the bucket with an IAM allowing full permissions from your compute instance. Additionally, on GCS, we use signed URLs which requires a corresponding service key to generate. You will need to create a service key and store it in env/google_service_key.json. If your content is otherwise publicly accessible, you could theoretically modify the app to avoid storing content in a bucket, and rather just point the player to the original, publicly accessible content.
You can optionally run media ingest separately from the search UI. This would enable you to run the more expensive GPU-enabled media ingest on-demand, with the search UI running on a lesser (free?) compute tier. This documentation assumes a combined ingest and UI server.
On AWS, a g4dn.xlarge EC2 instance type with a single NVIDIA T4 Tensor Core is appropriate. On GCP, a n1-standard-4 (4 vCPU, 2 core, 15GB RAM) with a single Tesla T4 GPU is appropriate.
I found these instance types provide a good balance between cost (~ 0.50 per minute) and CPU/GPU power (processing ~1 hour of video in about 10 minutes). That said, this demo should run on any modern CUDA-powered environment. I created a 256GB root volume to contain the requisite ML models and temporary media.
The installation script assumes use of Ubuntu 20/22 LTS. On GCP, I used an image that had the nvidia drivers pre-installed.
cd /home/ubuntu
git clone https://github.com/ty-elastic/intheirownwords.git
Create a file named env.vars in the intheirownwords/env directory on the compute instance with the following environment variables (do not wrap the values in quotes):
#for media storage (one or the other)
AWS_S3_BUCKET=
GCP_GCS_BUCKET=
#for huggingface models
HF_TOKEN=
#for export to elasticsearch
ES_ENDPOINT=
ES_USER=
ES_PASS=
Create a file name users.yaml in the intheirownwords/auth directory on the compute instance. Follow the instructions here to prepare the file.
setup/ubuntu.sh will setup the required dependencies on your compute instance to run the demo using docker containers and compose. setup/es.sh will setup requisite dependencies within Elasticsearch.
cd intheirownwords/setup
./ubuntu.sh
./es.sh
If you intend to further develop this demo, you can run ./ubuntu.sh -d true which will install the tools required to run directly on the host.
After install, and before use, please logout and log back into your compute instance to ensure docker is available without use of sudo.
We use docker compose for orchestration. From intheirownwords, run docker-compose up --detach.