Build LLM-powered applications using LangChain

This page introduces how to build LLM-powered applications using LangChain. The overviews on this page link to procedure guides in GitHub.

What is LangChain?

LangChain is an LLM orchestration framework that helps developers build generative AI applications or retrieval-augmented generation (RAG) workflows. It provides the structure, tools, and components to streamline complex LLM workflows.

For more information about LangChain, see the Google LangChain page. For more information about the LangChain framework, see the LangChain product documentation.

LangChain components for AlloyDB

Learn how to use LangChain with the LangChain Quickstart for AlloyDB. This quickstart creates an application that accesses a Netflix Movie dataset so that users can interact with movie data.

Vector store for AlloyDB

Vector store retrieves and stores documents and metadata from a vector database. Vector store gives an application the ability to perform semantic searches that interpret the meaning of a user query. This type of search is a called a vector search, and it can find topics that match the query conceptually. At query time, vector store retrieves the embedding vectors that are most similar to the embedding of the search request. In LangChain, a vector store takes care of storing embedded data and performing the vector search for you.

To work with vector store in AlloyDB, use the AlloyDBVectorStore class.

For more information, see the LangChain vector stores product documentation.

Vector store procedure guide

The AlloyDB guide for vector store shows you how to do the following:

  • Install the integration package and LangChain
  • Create an AlloyDBEngine object and configure a connection pool to your AlloyDB database
  • Initialize a table for the vector store
  • Set up an embedding service using VertexAIEmbeddings
  • Initialize AlloyDBVectorStore
  • Add and delete documents
  • Search for similar Documents
  • Add a vector index to improve search performance
  • Create a custom vector store to connect to a pre-existing AlloyDB for PostgreSQL database that has a table with vector embeddings

Document loader for AlloyDB

The document loader saves, loads, and deletes a LangChain Document objects. For example, you can load data for processing into embeddings and either store it in vector store or use it as a tool to provide specific context to chains.

To load documents from AlloyDB, use the AlloyDBLoader class. AlloyDBLoader returns a list of documents from a table using the first column for page content and all other columns for metadata. The default table has the first column as page content and the second column as JSON metadata. Each row becomes a document. Instructions for customizing these settings are in the procedure guide.

Use the AlloyDBSaver class to save and delete documents.

For more information, see the LangChain Document loaders topic.

Document loader procedure guide

The AlloyDB guide for document loader shows you how to do the following:

  • Install the integration package and LangChain
  • Load documents from a table
  • Add a filter to the loader
  • Customize the connection and authentication
  • Customize Document construction by specifying customer content and metadata
  • How to use and customize a AlloyDBSaver to store and delete documents

Chat message history for AlloyDB

Question and answer applications require a history of the things said in the conversation to give the application context for answering further questions from the user. The LangChain ChatMessageHistory class lets the application save messages to a database and retrieve them when needed to formulate further answers. A message can be a question, an answer, a statement, a greeting or any other piece of text that the user or application gives during the conversation. ChatMessageHistory stores each message and chains messages together for each conversation.

AlloyDB extends this class with AlloyDBChatMessageHistory.

Chat message history procedure guide

The AlloyDB guide for chat message history shows you how to do the following:

  • Install the integration package and LangChain
  • Create an AlloyDBEngine object and configure a connection pool to your AlloyDB database
  • Initialize a table
  • Initialize the AlloyDBChatMessageHistory class to add and delete messages
  • Create a chain for message history using the LangChain Expression Language (LCEL)