Saving pipelines as YAML

dumps writes a pipeline's components, settings and connections as YAML; loads rebuilds it. Loading refuses classes from modules you have not allowed.

yaml_text = pipeline.dumps()
print(yaml_text)

Each component is saved as its class path and its __init__ parameters, and connections as sender and receiver socket names. Custom components get this without extra code when their __init__ arguments are simple values.

saved = rag.dumps()
loaded = Pipeline.loads(saved)

Rebuilding a pipeline imports and creates the classes named in the file, so a YAML file from someone else could run any code. Haystack only loads classes from its own modules unless you allow more. shop_chat is your module, so name it:

loaded = Pipeline.loads(rag.dumps(), allowed_modules=["shop_chat"])
print(list(loaded.graph.nodes))
question = "When do parcels ship?"
print(loaded.run({"retriever": {"query": question}, "prompt": {"question": question}, "answer": {"query": question}})["answer"]["answers"][0].data)

The loaded pipeline has the same components and answers the same way. The documents are not in the YAML: the retriever's store is saved as its settings and its index id. An InMemoryDocumentStore keeps documents in memory shared by index for the whole process, so the loaded store found them. In a new process the same YAML gives an empty store; a database-backed store would still have them.