config

Configurations

This folder contains pipelines and individual components configurations files needed to for creating pipelines for inference and benchmarking.

Full Pipelines

There are several ready-to-run full pipelines. They are compatible with the haystack pipeline API.

• qa_pipeline - Elasticsearch retriever, SBERT cross encoder reranker and an FiD reader.
• qa_plaid- ColBERT/PLAID retriever and an FiD reader.
• summarization_pipeline_long_t5 - Elasticsearch retriever, SBERT cross encoder reranker and a large input LongT5 reader.
• summarization_pipeline_flan_t5 - Elasticsearch retriever, SBERT cross encoder reranker and an instruction tuned FLAN-T5 reader.
• qa_diffusion_pipeline - Elasticsearch retriever, SBERT cross encoder reranker, an FiD reader and a stable diffusion image generation reader.

Components Configuration

The following folders: data, embedder, image_generator, reader, reranker, retriever and store contain components' configuration YAML files. Those can be adapted to use specific models by providing a local path or a HuggingFace hub's model name or dataset name. After components are chosen, one needs to run the Pipeline Generation script in order to generate a pipeline configuration which describes the components and how they are connected. For example:

python generate_pipeline.py --path "retriever,reranker,reader"  \
       --store config/store/elastic.yaml                        \
       --retriever config/retriever/bm25.yaml                   \
       --reranker config/reranker/sbert.yaml                    \
       --reader config/reader/FiD.yaml                          \
       --file pipeline.yaml

This means you want to create a pipeline containing an elasticsearch based data store, a retriever based on BM25 algorithm, an SBERT cross encoding reranker and an FiD reader pipeline. The pipeline is saved as pipeline.yaml and can be used to create a web server to run queries:

python -m fastrag.rest_api.application pipeline.yaml

See Running Pipelines section for more details.