Skip to content

Latest commit

 

History

History
80 lines (58 loc) · 3.8 KB

LangChain-Zechang-20230523.md

File metadata and controls

80 lines (58 loc) · 3.8 KB

How to develop applications with LangChain?

Author: Zechang Sun

Date: May 23, 2023

for Data Science Club in DoA, Tsinghua

Ref: https://python.langchain.com/en/latest

What is LangChain?

LangChain is a framework for developing applications powered by language models. We believe that the most powerful and differentiated applications will not only call out to a language model, but will also be:

  1. Data-aware: connect a language model to other sources of data;
  2. Agents: allow a language model to interact with its environment.

Key Components of LangChain

  • Models

    • LLMs: large language models, which take a text string as input, and return a text string as output;
    • Chat Models: usually backed by a language model, but their APIs are more structured, these models take a list of Chat Messages as input, and return a Chat Message;
    • Text Embedding Models: take text as input and return a list of floats.

  • Prompts

    • PromptValue: an input to a model;

    • Promapt Templates: in charge of constructing a PromptValue;

    • Example Selectors: hardcoded or dynamically selected examples in prompts;

    • Output Parsers: get more structured information than just text back, instruct the model how output should be formatted and parse output into the desired format if necessary.

  • Indexes: how to structure documents so that LLMs can best interact with them

    • Document Loaders: classes responsible for loading documents from various sources;
    • Text Splitters: classes responsible for splitting text into smaller chunks;
    • VectorStores: the most common type of index, which relies on embeddings;
    • Retrievers: interface for fetching relevant documents to combine with language models.

  • Memory: store and retrieve data in the process of a conversation.

    • Two main methods:

      • Based on the input, fetch any relevant pieces of data;
      • Based on the input and output, update state accordingly.

    • Two main types of memory:

      • Short term: how to pass data in the context of a singluar conversion;
      • Long term: how to fetch and update information between conversations.

  • Chain: an incredibly generic concept which returns to a sequence of modular components (or other chains), combined in a particular way to accomplish a common use case;

  • Agents: has access to a suite of tools, depending on the user input, the agent can decide which, if any, of these tools to call.

Key Concept for developing applications

  • Chain of Thought
  • encourage the model to generate a series of intermediate reasoning steps
  • simply include "Let's think step-by-step" in the prompt
  • Action Plan Generation
  • use a language model to generate actions to take
  • results of these actions can then be fed back into the language model to generate a subsequent action
  • ReAct
    • combine Chain-of-Thought prompting with action plan generation
    • induce the model to think about what action to take, then take it
  • Self-ask
  • the model explicitly ask itself follow-up questions, which are then answered by an external search engine
  • Prompt Chaining
    • combing multiple LLM calls, with the output of one-step being the input to the next
  • Memetic Proxy
    • encourage the LLm to respond in a certain way framing the discussion in a context that the model knows of and that will result in that type of response
  • Self Consistency
    • sample a diverse set of reasoning paths and then selects the most consistent answer
    • is most effective when combined with Chain-of-thought prompting
  • Inception
    • also called First Person Instruction
    • encourage the model to think a certain way by including the start of the model's response in the prompt
  • MemPrompt
    • maintains a memory of errors and user feedback
    • use them to prevent repetition of mistakes