langchain/docs/extras/integrations/vectorstores/zep.ipynb

Zep

Zep is an open source long-term memory store for LLM applications. Zep makes it easy to add relevant documents, chat history memory & rich user data to your LLM app's prompts.

Note: The ZepVectorStore works with Documents and is intended to be used as a Retriever. It offers separate functionality to Zep's ZepMemory class, which is designed for persisting, enriching and searching your user's chat history.

Why Zep's VectorStore? 🤖🚀

Zep automatically embeds documents added to the Zep Vector Store using low-latency models local to the Zep server. The Zep client also offers async interfaces for all document operations. These two together with Zep's chat memory functionality make Zep ideal for building conversational LLM apps where latency and performance are important.

Installation

Follow the Zep Quickstart Guide to install and get started with Zep.

Usage

You'll need your Zep API URL and optionally an API key to use the Zep VectorStore. See the Zep docs for more information.

In the examples below, we're using Zep's auto-embedding feature which automatically embed documents on the Zep server using low-latency embedding models.

Note

• These examples use Zep's async interfaces. Call sync interfaces by removing the a prefix from the method names.
• If you pass in an Embeddings instance Zep will use this to embed documents rather than auto-embed them. You must also set your document collection to isAutoEmbedded === false.
• If you set your collection to isAutoEmbedded === false, you must pass in an Embeddings instance.

Load or create a Collection from documents

In [1]:

from uuid import uuid4

from langchain.document_loaders import WebBaseLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores import ZepVectorStore
from langchain.vectorstores.zep import CollectionConfig

ZEP_API_URL = "http://localhost:8000"  # this is the API url of your Zep instance
ZEP_API_KEY = "<optional_key>"  # optional API Key for your Zep instance
collection_name = f"babbage{uuid4().hex}"  # a unique collection name. alphanum only

# Collection config is needed if we're creating a new Zep Collection
config = CollectionConfig(
    name=collection_name,
    description="<optional description>",
    metadata={"optional_metadata": "associated with the collection"},
    is_auto_embedded=True,  # we'll have Zep embed our documents using its low-latency embedder
    embedding_dimensions=1536  # this should match the model you've configured Zep to use.
)

# load the document
article_url = "https://www.gutenberg.org/cache/epub/71292/pg71292.txt"
loader = WebBaseLoader(article_url)
documents = loader.load()

# split it into chunks
text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=0)
docs = text_splitter.split_documents(documents)

# Instantiate the VectorStore. Since the collection does not already exist in Zep,
# it will be created and populated with the documents we pass in.
vs = ZepVectorStore.from_documents(docs,
                                   collection_name=collection_name,
                                   config=config,
                                   api_url=ZEP_API_URL,
                                   api_key=ZEP_API_KEY
                                   )

In [2]:

# wait for the collection embedding to complete

async def wait_for_ready(collection_name: str) -> None:
    from zep_python import ZepClient
    import time

    client = ZepClient(ZEP_API_URL, ZEP_API_KEY)

    while True:
        c = await client.document.aget_collection(collection_name)
        print(
            "Embedding status: "
            f"{c.document_embedded_count}/{c.document_count} documents embedded"
        )
        time.sleep(1)
        if c.status == "ready":
            break


await wait_for_ready(collection_name)

Embedding status: 0/402 documents embedded
Embedding status: 0/402 documents embedded
Embedding status: 402/402 documents embedded

Simarility Search Query over the Collection

In [3]:

# query it
query = "what is the structure of our solar system?"
docs_scores = await vs.asimilarity_search_with_relevance_scores(query, k=3)

# print results
for d, s in docs_scores:
    print(d.page_content, " -> ", s, "\n====\n")

Tables necessary to determine the places of the planets are not less
necessary than those for the sun, moon, and stars. Some notion of the
number and complexity of these tables may be formed, when we state that
the positions of the two principal planets, (and these are the most
necessary for the navigator,) Jupiter and Saturn, require each not less
than one hundred and sixteen tables. Yet it is not only necessary to
predict the position of these bodies, but it is likewise expedient to  ->  0.8998482592744614 
====

tabulate the motions of the four satellites of Jupiter, to predict the
exact times at which they enter his shadow, and at which their shadows
cross his disc, as well as the times at which they are interposed
between him and the Earth, and he between them and the Earth.

Among the extensive classes of tables here enumerated, there are several
which are in their nature permanent and unalterable, and would never
require to be recomputed, if they could once be computed with perfect  ->  0.8976143854195493 
====

the scheme of notation thus applied, immediately suggested the
advantages which must attend it as an instrument for expressing the
structure, operation, and circulation of the animal system; and we
entertain no doubt of its adequacy for that purpose. Not only the
mechanical connexion of the solid members of the bodies of men and
animals, but likewise the structure and operation of the softer parts,
including the muscles, integuments, membranes, &c. the nature, motion,  ->  0.889982614061763 
====

Search over Collection Re-ranked by MMR

In [4]:

query = "what is the structure of our solar system?"
docs = await vs.asearch(query, search_type="mmr", k=3)

for d in docs:
    print(d.page_content, "\n====\n")

Tables necessary to determine the places of the planets are not less
necessary than those for the sun, moon, and stars. Some notion of the
number and complexity of these tables may be formed, when we state that
the positions of the two principal planets, (and these the most
necessary for the navigator,) Jupiter and Saturn, require each not less
than one hundred and sixteen tables. Yet it is not only necessary to
predict the position of these bodies, but it is likewise expedient to 
====

the scheme of notation thus applied, immediately suggested the
advantages which must attend it as an instrument for expressing the
structure, operation, and circulation of the animal system; and we
entertain no doubt of its adequacy for that purpose. Not only the
mechanical connexion of the solid members of the bodies of men and
animals, but likewise the structure and operation of the softer parts,
including the muscles, integuments, membranes, &c. the nature, motion, 
====

tabulate the motions of the four satellites of Jupiter, to predict the
exact times at which they enter his shadow, and at which their shadows
cross his disc, as well as the times at which they are interposed
between him and the Earth, and he between them and the Earth.

Among the extensive classes of tables here enumerated, there are several
which are in their nature permanent and unalterable, and would never
require to be recomputed, if they could once be computed with perfect 
====

Filter by Metadata

Use a metadata filter to narrow down results. First, load another book: "Adventures of Sherlock Holmes"

In [5]:

# Let's add more content to the existing Collection
article_url = "https://www.gutenberg.org/files/48320/48320-0.txt"
loader = WebBaseLoader(article_url)
documents = loader.load()

# split it into chunks
text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=0)
docs = text_splitter.split_documents(documents)

await vs.aadd_documents(docs)

await wait_for_ready(collection_name)

Embedding status: 402/1692 documents embedded
Embedding status: 402/1692 documents embedded
Embedding status: 552/1692 documents embedded
Embedding status: 702/1692 documents embedded
Embedding status: 1002/1692 documents embedded
Embedding status: 1002/1692 documents embedded
Embedding status: 1152/1692 documents embedded
Embedding status: 1302/1692 documents embedded
Embedding status: 1452/1692 documents embedded
Embedding status: 1602/1692 documents embedded
Embedding status: 1692/1692 documents embedded

We see results from both books. Note the source metadata

In [6]:

query = "Was he interested in astronomy?"
docs = await vs.asearch(query, search_type="similarity", k=3)

for d in docs:
    print(d.page_content, " -> ", d.metadata, "\n====\n")

by that body to Mr Babbage:--'In no department of science, or of the
arts, does this discovery promise to be so eminently useful as in that
of astronomy, and its kindred sciences, with the various arts dependent
on them. In none are computations more operose than those which
astronomy in particular requires;--in none are preparatory facilities
more needful;--in none is error more detrimental. The practical
astronomer is interrupted in his pursuit, and diverted from his task of  ->  {'source': 'https://www.gutenberg.org/cache/epub/71292/pg71292.txt'} 
====

possess all knowledge which is likely to be useful to him in his work,
and this I have endeavored in my case to do. If I remember rightly, you
on one occasion, in the early days of our friendship, defined my limits
in a very precise fashion.”

“Yes,” I answered, laughing. “It was a singular document. Philosophy,
astronomy, and politics were marked at zero, I remember. Botany
variable, geology profound as regards the mud-stains from any region  ->  {'source': 'https://www.gutenberg.org/files/48320/48320-0.txt'} 
====

in all its relations; but above all, with Astronomy and Navigation. So
important have they been considered, that in many instances large sums
have been appropriated by the most enlightened nations in the production
of them; and yet so numerous and insurmountable have been the
difficulties attending the attainment of this end, that after all, even
navigators, putting aside every other department of art and science,
have, until very recently, been scantily and imperfectly supplied with  ->  {'source': 'https://www.gutenberg.org/cache/epub/71292/pg71292.txt'} 
====

Let's try again using a filter for only the Sherlock Holmes document.

In [7]:

filter = {
    "where": {"jsonpath": "$[*] ? (@.source == 'https://www.gutenberg.org/files/48320/48320-0.txt')"},
}

docs = await vs.asearch(query, search_type="similarity", metadata=filter, k=3)

for d in docs:
    print(d.page_content, " -> ", d.metadata, "\n====\n")

possess all knowledge which is likely to be useful to him in his work,
and this I have endeavored in my case to do. If I remember rightly, you
on one occasion, in the early days of our friendship, defined my limits
in a very precise fashion.”

“Yes,” I answered, laughing. “It was a singular document. Philosophy,
astronomy, and politics were marked at zero, I remember. Botany
variable, geology profound as regards the mud-stains from any region  ->  {'source': 'https://www.gutenberg.org/files/48320/48320-0.txt'} 
====

the light shining upon his strong-set aquiline features. So he sat as I
dropped off to sleep, and so he sat when a sudden ejaculation caused me
to wake up, and I found the summer sun shining into the apartment. The
pipe was still between his lips, the smoke still curled upward, and the
room was full of a dense tobacco haze, but nothing remained of the heap
of shag which I had seen upon the previous night.

“Awake, Watson?” he asked.

“Yes.”

“Game for a morning drive?”  ->  {'source': 'https://www.gutenberg.org/files/48320/48320-0.txt'} 
====

“I glanced at the books upon the table, and in spite of my ignorance
of German I could see that two of them were treatises on science, the
others being volumes of poetry. Then I walked across to the window,
hoping that I might catch some glimpse of the country-side, but an oak
shutter, heavily barred, was folded across it. It was a wonderfully
silent house. There was an old clock ticking loudly somewhere in the
passage, but otherwise everything was deadly still. A vague feeling of  ->  {'source': 'https://www.gutenberg.org/files/48320/48320-0.txt'} 
====