LangChain: Chat with Your Data

Overview

LangChain is a framework designed to enable interaction with data using language models (LLMs). It leverages techniques such as Retrieval Augmented Generation (RAG) for efficient data integration.

Retrieval Augmented Generation (RAG)

RAG is a paradigm to enhance data utilization:

• Steps:
  1. Retrieve relevant documents.
  2. Load them into the context window or "working memory."
• Supported Data Sources:
  • PDFs
  • URLs
  • Databases
  • Notion

---

Document Loading and Splitting

Document Loaders

Loaders handle data access and format conversion:

• Supported Sources:
  • Websites
  • Databases
  • YouTube
  • arXiv
• Supported Data Formats:
  • PDF
  • HTML
  • JSON
  • Word
  • PowerPoint
• Output: Returns a list of Document objects.

Document Splitting

Documents are split into smaller, meaningful chunks while retaining context.

Example Code

from langchain.text_splitter import CharacterTextSplitter

splitter = CharacterTextSplitter(
    separator="\n\n",
    chunk_size=4000,
    chunk_overlap=200
)

Types of Splitters

• CharacterTextSplitter: Splits text by characters.
• MarkdownHeaderTextSplitter: Splits based on markdown headers.
• TokenTextSplitter: Splits text based on tokens.
• NLTKTextSplitter: Uses NLTK to split text into sentences.
• SpacyTextSplitter: Uses Spacy for sentence splitting.
• RecursiveCharacterTextSplitter: Attempts splitting by various characters.
• Language-Specific Splitters: Supports languages like Python, Markdown, etc.

---

Vector Stores and Embeddings

Vector Stores

• Store document splits and their embeddings.
• Enable efficient similarity search and retrieval.

Embeddings

• Text is converted into numerical vectors to capture semantic meaning.
• Similar texts yield similar vector representations.

---

Retrieval

Methods

• Basic Semantic Similarity: Matches queries with related content.
• Maximum Marginal Relevance (MMR): Ensures diverse responses.
• Metadata-Based Queries: Uses metadata for filtering results.
• LLM-Aided Retrieval:
  • Converts user questions into precise queries using LLMs.
  • Example: Self-Query for automatic query refinement.

Compression

• Shrinks retrieved responses to fit within LLM context by retaining only relevant information.

---

Question Answering

Steps

1. Retrieve relevant documents from the vector store.
2. Optionally compress the results to fit into the LLM context.
3. Pass the compressed results and the query to an LLM for the final answer.

RetrievalQA Chain

A chain combines retrieval and LLM-based processing:

• RetrievalQA.from_chain_type() supports different methods:
  • Stuff: Includes all retrieved documents in the prompt.
  • Map-Reduce: Summarizes documents into key points.

---

Agents

Agents use LLMs to determine:

• What actions to take.
• The sequence of those actions.

Key Components

• PromptTemplate: Constructs prompts based on user input.
• Language Model: Processes the prompt and generates output.
• Output Parser: Converts model output into actionable data.

Example Agent Initialization

from langchain.agents import initialize_agent, AgentType

agent = initialize_agent(
    tools,
    llm,
    agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
    verbose=True
)