Key Concepts Of Langchain

1 minute read

LangChain contains common abstractions that are useful to build language applications. This post introduces some of those concepts.

1. Language Models

Language models (LLM) are the core of LangChain applications.
Their main function is text generation: given a prompt, generate a text that follows the prompt.
At the core of a language model is a neural network that takes a sequence of tokens as input and outputs a probability distribution over the next token. The probability distribution is learned from a large corpus of text.

Prompts are inputs passed to models to get desired outputs.
ChatPromptTemplate enable create template so that prompts can be generated automatically and reused repeatedly.

Parsers extract structured information from model outputs -> enable chaining of models.
ResponseSchema and StructuredOutputParser works together to create format instruction and parse model outputs.

Memory store and manage the context and history of the ongoing conversation, and allow the LLM to refer back to earlier parts of the conversation to generate relevant and context-aware responses.
ConversationBufferMemory stores the entire conversation history in a buffer, can become long and costly - O(n^2) for n turns.
ConversationBufferWindowMemory and ConversationTokenBufferMemory limit the size of the buffer to the last k turns or tokens. - O(n) for n turns.
ConversationSummaryBufferMemory uses an LLM to generate a summary of the conversation history - O(n) for n turns.

Chain is a fundamental concept in LangChain, it allows you to:

Combine an LLM with a prompt to create a functional unit that processes input.
Chain multiple subchains together to perform a sequence of operations on the input.
Route input data to different subchains based on conditions or requirements.
Easily switch between different types of chains to achieve various tasks.

Types of Chains:

LLMChain = LLM + a single prompt, the most basic type of chain.
SimpleSequentialChain and SequentialChain = a sequence of subchains.
LLMRouterChain used to route input data to different subchains based on specific conditions.

Embedding capture the semantic meaning of the text and allow comparisons in a vector space
VectorStoreIndexCreator create a vector store index. It helps create a vector database that stores vector representations generated by embeddings.
The two combined to enable information retrieval and answer questions from external knowledge bases.