Semantic Poem Search Demo using GNES
This demo shows you how to build a semantic poem search using GNES.
Prerequisite
You need to have Docker installed on your machine.
As a start, run the following command to initialize the Docker Swarm environment.
docker swarm init1. Build images from local
make buildYou can also build DistilBert or Flair encoder by using make build encoder=distilbert- or make build encoder=flair-, respectively. Note the - at the end.
(Optional - Not yet available on dockerhub) Download pre-built images
If you have trouble building images yourself (e.g. due to slow connection or proxy issues), you can also pull our pre-build images for this demo. However, they may be outdated.
make pull2. Index all poems
# start the index server stack
make index
# wait for a minute until all services started and grpc is ready
# indexing 10 documents
make client_index d=10
3. Query and show top-10 poems
# start the query server stack
make query
# wait for a minute until all services started and grpc is ready
make client_query
Other commands
| Command | Description |
|---|---|
make clean |
Remove all docker stacks and release all network resources |
make clean_data |
Remove all index files in .cache |
make deploy_query |
Deploy the stack described in demo-poem-query.yml for query |
make deploy_index |
Deploy the stack described in demo-poem-index.yml for index |
Next steps
There are many ways to improve the quality of search results. To name a few,
- Score function: right now the demo is using the average of all chunks' scores as the final relevance score. This can be improved by using more sophisticated scoring functions as listed here.
- Encoder: instead of using pytorch-transformers or Flair, you can customize or build your new encoder pipeline. See how we build our own Flair in this demo and import it into GNES using
--py_path. - Preprocessor: this demo split a document into chunks using punctuations (the default behavior of
SentSplitPreprocessor), you may customize it as well.
If you find any interesting results, please feel free to contribute.
We are considering using this demo as a part of integration test and add it to the CICD pipeline of GNES master. So we may keep the master branch of this demo as simple as possible
Troubleshooting
Failed to remove some resources from stack: my-gnes
If you encounter the following errors when doing make, simply wait for couple of seconds. Docker is sometimes slow on recycling network device.
rm -rf .cache && mkdir .cache && docker stack rm my-gnes
Removing network my-gnes_default
Failed to remove network b0ei205abak98pn84yj9f6u70: Error response from daemon: network b0ei205abak98pn84yj9f6u70 not foundFailed to remove some resources from stack: my-gnes
make: *** [clean] Error 1Creating service my-gnes_Router40
failed to create service my-gnes_Router40: Error response from daemon: network my-gnes_default not found
make: *** [deploy] Error 1Force rebuilding images after code change
Thanks to the Docker cache, local change on the code may not be directly reflected in the Docker image. Therefore you can find placeholder such as RUN echo 5 in some Dockerfile. Increase this number will trigger make build to rebuild the image accordingly.
Be careful on the version mismatch
To update the local GNES image, please do docker pull gnes/gnes:latest-alpine before make build.