LangChain
Integration of YDB with LangChain allows using YDB as a vector store for RAG applications.
This integration allows developers to efficiently manage, query, and retrieve vectorized data, which is the foundation for modern applications related to natural language processing, search, and data analysis. Using embedding models, users can create sophisticated systems capable of understanding and retrieving information based on semantic similarity.
The integration is available for Python and JavaScript.
Installation
To use this integration, install a local YDB. For more information, see Install and start YDB.
Also install the LangChain packages and an embedding model for the required language:
Python
JavaScript
pip install -qU langchain-ydb
pip install -qU langchain-huggingface
npm install @ydbjs/langchain @langchain/core
npm install @langchain/community @huggingface/transformers
Initialization
To create a vector store YDB, you need to specify an embedding model and connection parameters:
Python
JavaScript
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_ydb.vectorstores import YDB, YDBSearchStrategy, YDBSettings
embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-mpnet-base-v2")
settings = YDBSettings(
host="localhost",
port=2136,
database="/local",
table="ydb_example",
strategy=YDBSearchStrategy.COSINE_SIMILARITY,
)
vector_store = YDB(embeddings, config=settings)
import { HuggingFaceTransformersEmbeddings } from "@langchain/community/embeddings/huggingface_transformers";
import { YDBSearchStrategy, YDBVectorStore } from "@ydbjs/langchain";
const embeddings = new HuggingFaceTransformersEmbeddings({
model: "sentence-transformers/all-mpnet-base-v2",
});
const vectorStore = new YDBVectorStore(embeddings, {
connectionString: "grpc://localhost:2136/local",
table: "ydb_example",
strategy: YDBSearchStrategy.CosineSimilarity,
});
Managing the Vector Store
Once you have created a vector store, you can interact with it by adding and removing various items.
Adding Items
Prepare the documents for processing:
Python
JavaScript
from uuid import uuid4
from langchain_core.documents import Document
uuids = [str(uuid4()) for _ in range(10)]
documents = [
Document(
page_content="I had chocolate chip pancakes and scrambled eggs for breakfast this morning.",
metadata={"source": "tweet"},
id=uuids[0],
),
Document(
page_content="The weather forecast for tomorrow is cloudy and overcast, with a high of 62 degrees.",
metadata={"source": "news"},
id=uuids[1],
),
Document(
page_content="Building an exciting new project with LangChain - come check it out!",
metadata={"source": "tweet"},
id=uuids[2],
),
Document(
page_content="Robbers broke into the city bank and stole $1 million in cash.",
metadata={"source": "news"},
id=uuids[3],
),
Document(
page_content="Wow! That was an amazing movie. I can't wait to see it again.",
metadata={"source": "tweet"},
id=uuids[4],
),
Document(
page_content="Is the new iPhone worth the price? Read this review to find out.",
metadata={"source": "website"},
id=uuids[5],
),
Document(
page_content="The top 10 soccer players in the world right now.",
metadata={"source": "website"},
id=uuids[6],
),
Document(
page_content="LangGraph is the best framework for building stateful, agentic applications!",
metadata={"source": "tweet"},
id=uuids[7],
),
Document(
page_content="The stock market is down 500 points today due to fears of a recession.",
metadata={"source": "news"},
id=uuids[8],
),
Document(
page_content="I have a bad feeling I am going to get deleted :(",
metadata={"source": "tweet"},
id=uuids[9],
),
]
Add the documents to the vector store:
ids = vector_store.add_documents(documents=documents)
import { Document } from "@langchain/core/documents";
import { randomUUID } from "node:crypto";
const uuids = Array.from({ length: 10 }, () => randomUUID());
const documents = [
new Document({
pageContent: "I had chocolate chip pancakes and scrambled eggs for breakfast this morning.",
metadata: { source: "tweet" },
id: uuids[0],
}),
new Document({
pageContent: "The weather forecast for tomorrow is cloudy and overcast, with a high of 62 degrees.",
metadata: { source: "news" },
id: uuids[1],
}),
new Document({
pageContent: "Building an exciting new project with LangChain - come check it out!",
metadata: { source: "tweet" },
id: uuids[2],
}),
new Document({
pageContent: "Robbers broke into the city bank and stole $1 million in cash.",
metadata: { source: "news" },
id: uuids[3],
}),
new Document({
pageContent: "Wow! That was an amazing movie. I can't wait to see it again.",
metadata: { source: "tweet" },
id: uuids[4],
}),
new Document({
pageContent: "Is the new iPhone worth the price? Read this review to find out.",
metadata: { source: "website" },
id: uuids[5],
}),
new Document({
pageContent: "The top 10 soccer players in the world right now.",
metadata: { source: "website" },
id: uuids[6],
}),
new Document({
pageContent: "LangGraph is the best framework for building stateful, agentic applications!",
metadata: { source: "tweet" },
id: uuids[7],
}),
new Document({
pageContent: "The stock market is down 500 points today due to fears of a recession.",
metadata: { source: "news" },
id: uuids[8],
}),
new Document({
pageContent: "I have a bad feeling I am going to get deleted :(",
metadata: { source: "tweet" },
id: uuids[9],
}),
];
Add the documents to the vector store:
const ids = await vectorStore.addDocuments(documents);
Deleting Items
Items are deleted from the vector store by ID using the delete function:
Python
JavaScript
vector_store.delete(ids=[ids[-1]])
await vectorStore.delete({ ids: [ids.at(-1)] });
Querying the Vector Store
After creating the vector store and adding the necessary documents, you can perform search queries during chain or agent execution.
Direct Query
Similarity Search
A simple similarity search can be performed as follows:
Python
JavaScript
results = vector_store.similarity_search(
"LangChain provides abstractions to make working with LLMs easy",
k=2,
)
for res in results:
print(f"* {res.page_content} [{res.metadata}]")
const results = await vectorStore.similaritySearch(
"LangChain provides abstractions to make working with LLMs easy",
2,
);
for (const res of results) {
console.log(`* ${res.pageContent} [${JSON.stringify(res.metadata)}]`);
}
Result:
* Building an exciting new project with LangChain - come check it out! [{'source': 'tweet'}]
* LangGraph is the best framework for building stateful, agentic applications! [{'source': 'tweet'}]
Similarity Search with Score
You can also perform a search with a score:
Python
JavaScript
results = vector_store.similarity_search_with_score(
"Will it be hot tomorrow?",
k=3,
)
for res, score in results:
print(f"* [SIM={score:.3f}] {res.page_content} [{res.metadata}]")
const results = await vectorStore.similaritySearchWithScore(
"Will it be hot tomorrow?",
3,
);
for (const [res, score] of results) {
console.log(`* [SIM=${score.toFixed(3)}] ${res.pageContent} [${JSON.stringify(res.metadata)}]`);
}
Result:
* [SIM=0.595] The weather forecast for tomorrow is cloudy and overcast, with a high of 62 degrees. [{'source': 'news'}]
* [SIM=0.212] I had chocolate chip pancakes and scrambled eggs for breakfast this morning. [{'source': 'tweet'}]
* [SIM=0.118] Wow! That was an amazing movie. I can't wait to see it again. [{'source': 'tweet'}]
Filtering
Search using filters is performed as follows:
Python
JavaScript
results = vector_store.similarity_search_with_score(
"What did I eat for breakfast?",
k=4,
filter={"source": "tweet"},
)
for res, _ in results:
print(f"* {res.page_content} [{res.metadata}]")
const results = await vectorStore.similaritySearchWithScore(
"What did I eat for breakfast?",
4,
{ source: "tweet" },
);
for (const [res] of results) {
console.log(`* ${res.pageContent} [${JSON.stringify(res.metadata)}]`);
}
Result:
* I had chocolate chip pancakes and scrambled eggs for breakfast this morning. [{'source': 'tweet'}]
* Wow! That was an amazing movie. I can't wait to see it again. [{'source': 'tweet'}]
* Building an exciting new project with LangChain - come check it out! [{'source': 'tweet'}]
* LangGraph is the best framework for building stateful, agentic applications! [{'source': 'tweet'}]
Query via Retriever Transformation
The vector store can be transformed into a retriever for simplified use in chains.
An example is shown below:
Python
JavaScript
retriever = vector_store.as_retriever(
search_kwargs={
"k": 2,
"filter": {"source": "news"},
},
)
results = retriever.invoke("Stealing from the bank is a crime")
for res in results:
print(f"* {res.page_content} [{res.metadata}]")
const retriever = vectorStore.asRetriever({
k: 2,
filter: { source: "news" },
});
const results = await retriever.invoke("Stealing from the bank is a crime");
for (const res of results) {
console.log(`* ${res.pageContent} [${JSON.stringify(res.metadata)}]`);
}
Result:
* Robbers broke into the city bank and stole $1 million in cash. [{'source': 'news'}]
* The stock market is down 500 points today due to fears of a recession. [{'source': 'news'}]
Was the article helpful?
Previous