docs/tutorials/retrievers/

[giscus[bot]]

docs/tutorials/retrievers/

This tutorial will familiarize you with LangChain's document loader, embedding, and vector store abstractions. These abstractions are designed to support retrieval of data-- from (vector) databases and other sources-- for integration with LLM workflows. They are important for applications that fetch data to be reasoned over as part of model inference, as in the case of retrieval-augmented generation, or RAG (see our RAG tutorial here).

https://python.langchain.com/docs/tutorials/retrievers/

[alimoameri]

I get an error when setting Mistral API key like this (as suggested in the code):
os.environ["MISTRALAI_API_KEY"] = getpass.getpass("Enter API key for MistralAI: ")

The correct way is to set MISTRAL_API_KEY instead of MISTRALAI_API_KEY:
os.environ["MISTRAL_API_KEY"] = getpass.getpass("Enter API key for MistralAI:")

[sravan953]

I am trying out a simple vector similarity search and I am seeing poor performance across Nomic embeddings and Llama 32. 1b embeddings. Any help is appreciated.

Code to reproduce

from langchain_core.documents import Document
from langchain_core.vectorstores import InMemoryVectorStore
from langchain_ollama import OllamaEmbeddings
from langchain_text_splitters import RecursiveCharacterTextSplitter


def add_documents_to_vector_store(vector_store, documents):
    vector_store.add_documents(documents)


def main(text_document: str, question: str):
    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=256,
        chunk_overlap=64,
        add_start_index=True,
    )
    text_splits = text_splitter.split_text(text_document)
    arr_documents = [
        Document(
            page_content=split, metadata={"source": "wikipedia", "page_number": i + 1}
        )
        for i, split in enumerate(text_splits)
    ]

    # nomic_embeddings = NomicEmbeddings(
    #     model="nomic-embed-text-v1.5",
    #     # dimensionality=64,
    #     inference_mode="local",
    #     # device="gpu",
    # )
    nomic_embeddings = OllamaEmbeddings(model="llama3.2:1b")
    vec_store = InMemoryVectorStore(nomic_embeddings)
    add_documents_to_vector_store(vec_store, arr_documents)

    results = vec_store.similarity_search_with_score(question, k=1)[0]
    best_result, best_score = results
    print(best_result.page_content)
    print(best_result.metadata, best_score)


if __name__ == "__main__":
    text_document = """
    India, officially the Republic of India,[j][21] is a country in South Asia. It is the seventh-largest country by area; the most populous country from June 2023 onwards;[22][23] and since its independence in 1947, the world's most populous democracy.[24][25][26] Bounded by the Indian Ocean on the south, the Arabian Sea on the southwest, and the Bay of Bengal on the southeast, it shares land borders with Pakistan to the west;[k] China, Nepal, and Bhutan to the north; and Bangladesh and Myanmar to the east. In the Indian Ocean, India is near Sri Lanka and the Maldives; its Andaman and Nicobar Islands share a maritime border with Thailand, Myanmar, and Indonesia.

    Modern humans arrived on the Indian subcontinent from Africa no later than 55,000 years ago.[28][29][30] Their long occupation, predominantly in isolation as hunter-gatherers, has made the region highly diverse, second only to Africa in human genetic diversity.[31] Settled life emerged on the subcontinent in the western margins of the Indus river basin 9,000 years ago, evolving gradually into the Indus Valley Civilisation of the third millennium BCE.[32] By 1200 BCE, an archaic form of Sanskrit, an Indo-European language, had diffused into India from the northwest.[33][34] Its hymns recorded the dawning of Hinduism in India.[35] India's pre-existing Dravidian languages were supplanted in the northern regions.[36] By 400 BCE, caste had emerged within Hinduism,[37] and Buddhism and Jainism had arisen, proclaiming social orders unlinked to heredity.[38] Early political consolidations gave rise to the loose-knit Maurya and Gupta Empires.[39] Widespread creativity suffused this era,[40] but the status of women declined,[41] and untouchability became an organized belief.[l][42] In South India, the Middle kingdoms exported Dravidian language scripts and religious cultures to the kingdoms of Southeast Asia.[43]
    """
    question = "When did India gain independence?"

    main(text_document, question)

Output

the Maldives; its Andaman and Nicobar Islands share a maritime border with Thailand, Myanmar, and Indonesia.
{'source': 'wikipedia', 'page_number': 4} 0.5966474281799025

^ Above output is not relevant at all to the question prompt.

[dumbelldoor]

Hey

Could you try changing the embedding model?

[sravan953]

Actually, changing the vector store to FAISS or ChromaDB made a big difference.

[dumbelldoor]

You were trying to run the model locally. Ollama is usually suitable for running models locally.
Glad It's working for you.

[sravan953]

No model involved in this specific example, but yes, using the Ollama integration in general.

[ShahanAhmedDev]

They Voyage AI for import is wrong.
It should be:
from langchain_voyageai import VoyageAIEmbeddings

Currently its
from langchain-voyageai import VoyageAIEmbeddings.

Just some helpful tidbid, maybe the authors can quickly fix it aswell!

[irisfaustina]

Forth the embeddings example, the model "text-embedding-3-large" by default returns a vector of length 3072 not 1536.

[alainkaiser]

true :) also noted that one!

[lroberts]

hi
i'm new to langchain and llms in general, hoping someone can point me in the right direction.... i'm not getting the good results as shown in this tutorial.

i have tried running sravan953's code and i also get (the same) poor results ("the Maldives; its Andaman and Nicobar Islands share a maritime border..."), but unlike sravan, even when i switch to faiss (i did not try chromadb) it still gives poor results (the exact same results). the top 3 look like this:
RESULT:
the Maldives; its Andaman and Nicobar Islands share a maritime border with Thailand, Myanmar, and Indonesia.
{'source': 'wikipedia', 'page_number': 4} 1.226837

RESULT:
gradually into the Indus Valley Civilisation of the third millennium BCE.[32] By 1200 BCE, an archaic form of Sanskrit, an Indo-European language, had diffused into India from the northwest.[33][34] Its hymns recorded the dawning of Hinduism in India.[35]
{'source': 'wikipedia', 'page_number': 7} 1.2655632

RESULT:
scripts and religious cultures to the kingdoms of Southeast Asia.[43]
{'source': 'wikipedia', 'page_number': 11} 1.2974433

i followed the tutorial code and it also has poor results:
from langchain_community.document_loaders import PyPDFLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_ollama import OllamaEmbeddings
import faiss
from langchain_community.docstore.in_memory import InMemoryDocstore
from langchain_community.vectorstores import FAISS

file_path = "./nke-10k-2023.pdf"
loader = PyPDFLoader(file_path)
docs = loader.load()
print(len(docs))

text_splitter = RecursiveCharacterTextSplitter(
chunk_size=1000, chunk_overlap=200, add_start_index=True
)
all_splits = text_splitter.split_documents(docs)

embeddings = OllamaEmbeddings(model="llama3.2")
vector_1 = embeddings.embed_query(all_splits[0].page_content)
vector_2 = embeddings.embed_query(all_splits[1].page_content)

assert len(vector_1) == len(vector_2)
print(f"Generated vectors of length {len(vector_1)}\n")
print(vector_1[:10])

embedding_dim = len(vector_1)
index = faiss.IndexFlatL2(embedding_dim)

vector_store = FAISS(
embedding_function=embeddings,
index=index,
docstore=InMemoryDocstore(),
index_to_docstore_id={},
)

ids = vector_store.add_documents(documents=all_splits)

results = vector_store.similarity_search(
"How many distribution centers does Nike have in the US?"
)

print("How many distribution centers does Nike have in the US?")
print(results[0])

tried a bunch of other prompts. zeroshot was bad to terrible and oneshot did not seem to improve it noticeably. changing the chunk size and overlap down (250/100 and 100/40) also did not help.

is it just ollama is no good? how did sravan get it to be functional? any tips are greatly appreciated, i'm sure i'm doing something obvious wrong!