This code deploys a docker container with a question-answering system based on Haystack.
docker compose up -d This deploys all the necessary containers:
- Milvus to store documents and their embeddings.
- FastAPI endpoint
The main parameters are stored in settings.py:
DOC_DIR: directory where the raw txt files will be downloadedRETRIEVER_MODEL: HuggingFace Hub model name to retrieve relevant documents before performing QARETRIEVER_EMB_SIZE: embedding size of the RETRIEVER_MODELQA_MODEL: HuggingFace Hub model name to retrieve answersUSE_GPU: whether to use GPU or not
Also, you may want to select another PyTorch base image (by changing BASE_IMAGE in docker-compose.yml) but beware of dependencies breaks.
Other changes demand code modification. You can change:
- Document store (FAISS, ElasticSearch, and other), default is Milvus store
- Documents chunk sizes (see preprocessor in
download_data.py) - Retrieval method (BM25, TF-IDF), default is embeddings similarity retr 501E ieval
Modify download_data.py to populate the database with your documents in txt format.