Faiss

Facebook AI Similarity Search (Faiss) is a library for efficient similarity search and clustering of dense vectors. It contains algorithms that search in sets of vectors of any size, up to ones that possibly do not fit in RAM. It also contains supporting code for evaluation and parameter tuning.

Faiss documentation.

This notebook shows how to use functionality related to the FAISS vector database. It will show functionality specific to this integration. After going through, it may be useful to explore relevant use-case pages to learn how to use this vectorstore as part of a larger chain.

Setup

The integration lives in the langchain-community package. We also need to install the faiss package itself. We will also be using OpenAI for embeddings, so we need to install those requirements. We can install these with:

pip install -U langchain-community faiss-cpu langchain-openai tiktoken

Note that you can also install faiss-gpu if you want to use the GPU enabled version

Since we are using OpenAI, you will need an OpenAI API Key.

import getpass
import os

os.environ["OPENAI_API_KEY"] = getpass.getpass()

It’s also helpful (but not needed) to set up LangSmith for best-in-class observability

# os.environ["LANGCHAIN_TRACING_V2"] = "true"
# os.environ["LANGCHAIN_API_KEY"] = getpass.getpass()

Ingestion

Here, we ingest documents into the vectorstore

# Uncomment the following line if you need to initialize FAISS with no AVX2 optimization
# os.environ['FAISS_NO_AVX2'] = '1'

from langchain_community.document_loaders import TextLoader
from langchain_community.vectorstores import FAISS
from langchain_openai import OpenAIEmbeddings
from langchain_text_splitters import CharacterTextSplitter

loader = TextLoader("../../modules/state_of_the_union.txt")
documents = loader.load()
text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
docs = text_splitter.split_documents(documents)
embeddings = OpenAIEmbeddings()
db = FAISS.from_documents(docs, embeddings)
print(db.index.ntotal)

API Reference:

Querying

Now, we can query the vectorstore. There a few methods to do this. The most standard is to use similarity_search.

query = "What did the president say about Ketanji Brown Jackson"
docs = db.similarity_search(query)

print(docs[0].page_content)

Tonight. I call on the Senate to: Pass the Freedom to Vote Act. Pass the John Lewis Voting Rights Act. And while you’re at it, pass the Disclose Act so Americans can know who is funding our elections. 

Tonight, I’d like to honor someone who has dedicated his life to serve this country: Justice Stephen Breyer—an Army veteran, Constitutional scholar, and retiring Justice of the United States Supreme Court. Justice Breyer, thank you for your service. 

One of the most serious constitutional responsibilities a President has is nominating someone to serve on the United States Supreme Court. 

And I did that 4 days ago, when I nominated Circuit Court of Appeals Judge Ketanji Brown Jackson. One of our nation’s top legal minds, who will continue Justice Breyer’s legacy of excellence.

As a Retriever

We can also convert the vectorstore into a Retriever class. This allows us to easily use it in other LangChain methods, which largely work with retrievers

retriever = db.as_retriever()
docs = retriever.invoke(query)

print(docs[0].page_content)

Tonight. I call on the Senate to: Pass the Freedom to Vote Act. Pass the John Lewis Voting Rights Act. And while you’re at it, pass the Disclose Act so Americans can know who is funding our elections. 

Tonight, I’d like to honor someone who has dedicated his life to serve this country: Justice Stephen Breyer—an Army veteran, Constitutional scholar, and retiring Justice of the United States Supreme Court. Justice Breyer, thank you for your service. 

One of the most serious constitutional responsibilities a President has is nominating someone to serve on the United States Supreme Court. 

And I did that 4 days ago, when I nominated Circuit Court of Appeals Judge Ketanji Brown Jackson. One of our nation’s top legal minds, who will continue Justice Breyer’s legacy of excellence.

Similarity Search with score

There are some FAISS specific methods. One of them is similarity_search_with_score, which allows you to return not only the documents but also the distance score of the query to them. The returned distance score is L2 distance. Therefore, a lower score is better.

docs_and_scores = db.similarity_search_with_score(query)

docs_and_scores[0]

(Document(page_content='Tonight. I call on the Senate to: Pass the Freedom to Vote Act. Pass the John Lewis Voting Rights Act. And while you’re at it, pass the Disclose Act so Americans can know who is funding our elections. \n\nTonight, I’d like to honor someone who has dedicated his life to serve this country: Justice Stephen Breyer—an Army veteran, Constitutional scholar, and retiring Justice of the United States Supreme Court. Justice Breyer, thank you for your service. \n\nOne of the most serious constitutional responsibilities a President has is nominating someone to serve on the United States Supreme Court. \n\nAnd I did that 4 days ago, when I nominated Circuit Court of Appeals Judge Ketanji Brown Jackson. One of our nation’s top legal minds, who will continue Justice Breyer’s legacy of excellence.', metadata={'source': '../../modules/state_of_the_union.txt'}),
 0.36913747)

It is also possible to do a search for documents similar to a given embedding vector using similarity_search_by_vector which accepts an embedding vector as a parameter instead of a string.

embedding_vector = embeddings.embed_query(query)
docs_and_scores = db.similarity_search_by_vector(embedding_vector)

Saving and loading

You can also save and load a FAISS index. This is useful so you don’t have to recreate it everytime you use it.

db.save_local("faiss_index")

new_db = FAISS.load_local("faiss_index", embeddings)

docs = new_db.similarity_search(query)

docs[0]

Document(page_content='Tonight. I call on the Senate to: Pass the Freedom to Vote Act. Pass the John Lewis Voting Rights Act. And while you’re at it, pass the Disclose Act so Americans can know who is funding our elections. \n\nTonight, I’d like to honor someone who has dedicated his life to serve this country: Justice Stephen Breyer—an Army veteran, Constitutional scholar, and retiring Justice of the United States Supreme Court. Justice Breyer, thank you for your service. \n\nOne of the most serious constitutional responsibilities a President has is nominating someone to serve on the United States Supreme Court. \n\nAnd I did that 4 days ago, when I nominated Circuit Court of Appeals Judge Ketanji Brown Jackson. One of our nation’s top legal minds, who will continue Justice Breyer’s legacy of excellence.', metadata={'source': '../../../state_of_the_union.txt'})

Serializing and De-Serializing to bytes

you can pickle the FAISS Index by these functions. If you use embeddings model which is of 90 mb (sentence-transformers/all-MiniLM-L6-v2 or any other model), the resultant pickle size would be more than 90 mb. the size of the model is also included in the overall size. To overcome this, use the below functions. These functions only serializes FAISS index and size would be much lesser. this can be helpful if you wish to store the index in database like sql.

from langchain_community.embeddings.huggingface import HuggingFaceEmbeddings

pkl = db.serialize_to_bytes()  # serializes the faiss
embeddings = HuggingFaceEmbeddings(model_name="all-MiniLM-L6-v2")

db = FAISS.deserialize_from_bytes(
    embeddings=embeddings, serialized=pkl
)  # Load the index

API Reference:

HuggingFaceEmbeddings

Merging

You can also merge two FAISS vectorstores

db1 = FAISS.from_texts(["foo"], embeddings)
db2 = FAISS.from_texts(["bar"], embeddings)

db1.docstore._dict

db2.docstore._dict

{'807e0c63-13f6-4070-9774-5c6f0fbb9866': Document(page_content='bar', metadata={})}

db1.merge_from(db2)

db1.docstore._dict

{'068c473b-d420-487a-806b-fb0ccea7f711': Document(page_content='foo', metadata={}),
 '807e0c63-13f6-4070-9774-5c6f0fbb9866': Document(page_content='bar', metadata={})}

Similarity Search with filtering

FAISS vectorstore can also support filtering, since the FAISS does not natively support filtering we have to do it manually. This is done by first fetching more results than k and then filtering them. This filter is either a callble that takes as input a metadata dict and returns a bool, or a metadata dict where each missing key is ignored and each present k must be in a list of values. You can also set the fetch_k parameter when calling any search method to set how many documents you want to fetch before filtering. Here is a small example:

from langchain_core.documents import Document

list_of_documents = [
    Document(page_content="foo", metadata=dict(page=1)),
    Document(page_content="bar", metadata=dict(page=1)),
    Document(page_content="foo", metadata=dict(page=2)),
    Document(page_content="barbar", metadata=dict(page=2)),
    Document(page_content="foo", metadata=dict(page=3)),
    Document(page_content="bar burr", metadata=dict(page=3)),
    Document(page_content="foo", metadata=dict(page=4)),
    Document(page_content="bar bruh", metadata=dict(page=4)),
]
db = FAISS.from_documents(list_of_documents, embeddings)
results_with_scores = db.similarity_search_with_score("foo")
for doc, score in results_with_scores:
    print(f"Content: {doc.page_content}, Metadata: {doc.metadata}, Score: {score}")

API Reference:

Document

Content: foo, Metadata: {'page': 1}, Score: 5.159960813797904e-15
Content: foo, Metadata: {'page': 2}, Score: 5.159960813797904e-15
Content: foo, Metadata: {'page': 3}, Score: 5.159960813797904e-15
Content: foo, Metadata: {'page': 4}, Score: 5.159960813797904e-15

Now we make the same query call but we filter for only page = 1

results_with_scores = db.similarity_search_with_score("foo", filter=dict(page=1))
# Or with a callable:
# results_with_scores = db.similarity_search_with_score("foo", filter=lambda d: d["page"] == 1)
for doc, score in results_with_scores:
    print(f"Content: {doc.page_content}, Metadata: {doc.metadata}, Score: {score}")

Content: foo, Metadata: {'page': 1}, Score: 5.159960813797904e-15
Content: bar, Metadata: {'page': 1}, Score: 0.3131446838378906

Same thing can be done with the max_marginal_relevance_search as well.

results = db.max_marginal_relevance_search("foo", filter=dict(page=1))
for doc in results:
    print(f"Content: {doc.page_content}, Metadata: {doc.metadata}")

Content: foo, Metadata: {'page': 1}
Content: bar, Metadata: {'page': 1}

Here is an example of how to set fetch_k parameter when calling similarity_search. Usually you would want the fetch_k parameter >> k parameter. This is because the fetch_k parameter is the number of documents that will be fetched before filtering. If you set fetch_k to a low number, you might not get enough documents to filter from.

results = db.similarity_search("foo", filter=dict(page=1), k=1, fetch_k=4)
for doc in results:
    print(f"Content: {doc.page_content}, Metadata: {doc.metadata}")

Content: foo, Metadata: {'page': 1}

Delete

You can also delete records from vectorstore. In the example below db.index_to_docstore_id represents a dictionary with elements of the FAISS index.

print("count before:", db.index.ntotal)
db.delete([db.index_to_docstore_id[0]])
print("count after:", db.index.ntotal)

count before: 8
count after: 7

Setup​

Ingestion​

API Reference:

Querying​

As a Retriever​

Similarity Search with score​

Saving and loading​

Serializing and De-Serializing to bytes

API Reference:

Merging​

Similarity Search with filtering​

API Reference:

Delete​

Help us out by providing feedback on this documentation page:

Setup

Ingestion

Querying

As a Retriever

Similarity Search with score

Saving and loading

Merging

Similarity Search with filtering

Delete