Unable to load pipeline using load_from_yml containing FAISSDocumentStore node

[preethampaulose]

Describe the bug 'm trying to load the yml that was created using the save_to_yaml method. The pipeline was first tested to be working and saved. When I try to load the saved yaml file I'm getting error. The Pipeline consists of FaissDocumentStore, Retriever and Generator.

Error message could not determine a constructor for the tag 'tag:yaml.org,2002:python/object:faiss.swigfaiss.IndexFlat' in "myPipeline.haystack-pipeline.yml", line 5, column 18

Expected behavior Load the pipeline containing all the nodes

Additional context Add any other context about the problem here, like document types / preprocessing steps / settings of reader etc.

To Reproduce

1. Unzip the attached file
2. Run the "load saved faiss document.ipynb"

FAQ Check

• [x] Have you had a look at our new FAQ page?

System:

• OS: Windows 11
• GPU/CPU: CPU
• Haystack version (commit or version number): 1.12.2
• DocumentStore: FAISS
• Reader: none
• Retriever: DensePassageRetriever
• Generator: Seq2SeqGenerator

Attachment load_issue.zip

[ZanSara]

Hey @preethampaulose, you're using a quite outdated Haystack version. Would you mind trying to upgrade it and let us know if the issue is still present in the latest version?

[anakin87]

While investigating on https://github.com/deepset-ai/haystack/issues/4351, I discovered that this issue is still present.

To reproduce

from haystack.pipelines import Pipeline
from haystack.document_stores import FAISSDocumentStore
from haystack.nodes import EmbeddingRetriever

ds = FAISSDocumentStore(faiss_index_path="/home/anakin87/apps/experiments/doc-search/index/my_faiss_index.faiss",
                        faiss_config_path="/home/anakin87/apps/experiments/doc-search/index/my_faiss_index.json")

retriever = EmbeddingRetriever(
    document_store=ds,
    embedding_model="sentence-transformers/msmarco-distilbert-base-tas-b",
    model_format="sentence_transformers",
)

pipe = Pipeline()
pipe.add_node(component=retriever, name="Retriever", inputs=["Query"])

pipe.save_to_yaml("mypipe.yml")

Generated YAML

components:
- name: FAISSDocumentStore
  params:
    faiss_config_path: /home/anakin87/apps/experiments/doc-search/index/my_faiss_index.json
    faiss_index: !!python/object:faiss.swigfaiss.IndexFlat
      this: !!binary |
      !!! very long binary string !!!
    faiss_index_path: /home/anakin87/apps/experiments/doc-search/index/my_faiss_index.faiss
  type: FAISSDocumentStore
- name: Retriever
  params:
    document_store: FAISSDocumentStore
    embedding_model: sentence-transformers/msmarco-distilbert-base-tas-b
    model_format: sentence_transformers
  type: EmbeddingRetriever
pipelines:
- name: query
  nodes:
  - inputs:
    - Query
    name: Retriever
version: 1.19.0rc0

The generated YAML contains a very long binary string (the faiss index). My very first idea to solve this issue is about skipping this field in the save_to_yaml method.

[TomAtGithub]

Hey, I experienced the same problem. I was able to workaround this problem by removing all params except faiss_config_path and faiss_index_path from the FAISSDocumentStore. This makes sense because the FAISSDocumentStore constructor does not allow other params except the two mentioned if faiss_config_path is set (see reference). It seems like the FAISSDocumentStore.save_to_yaml() function does not take care of this rule.

Here is a working yaml (it assumes that a FAISSDocumentStore was indexed and saved before to the ./faiss directory):

components:
- name: FAISSDocumentStore
  params:
    faiss_config_path: ./faiss/config.json
    faiss_index_path: ./faiss/index.faiss
  type: FAISSDocumentStore
- name: Retriever
  params:
    document_store: FAISSDocumentStore
    embedding_model: sentence-transformers/all-MiniLM-L6-v2
  type: EmbeddingRetriever
- name: Reader
  params:
    model_name_or_path: sentence-transformers/all-MiniLM-L6-v2
  type: FARMReader
pipelines:
- name: query
  nodes:
  - inputs:
    - Query
    name: Retriever
  - inputs:
    - Retriever
    name: Reader
version: 1.22.1

I am using version 1.22.1 and not the latest version because of bug #5749