LangChain for LLM development

Notes on the short course from DeepLearning.AI

General

LangChain (LC) provides abstraction for interacting with LLM APIs, i.e. prompt templates and output parsers
LLMs are stateless, but when building a chat interface, we can supply the entire conversation as context
Embeddings create numerical representation for pieces of text covering their content/meaning - text with similar content will have similar vectors
Vector databases contain embedding vectors and the corresponding chunk of text (which are usually smaller than the original document/text)
Querying a vector database means creating an embedding for the search query and retrieving n most-similar chunks of text
Chain-of-Thought Reasoning (ReAct)

Check out prompts for common operations
Chains and memory (especially ConversationSummaryBufferMemory) lend themselves to building chat interfaces
Use chains to use (multiple) output from one prompt as input to another prompt
Memory can also be stored and retrieved in/from external databases, i.e. vector databases
LangChain also supports developing Q&A applications over documents
Use QAGenerationChain to generate question/answer pairs automatically (using LLM)
Use QAEvalChain to evaluate question/answer pairs automatically (using LLM)
Use langchain.debug = True to get debugging information
Agents use LLMs as a reasoning engine and can use tools, i.e. a calculator or searching information on Wikipedia, to complete tasks
You can define your own tools by writing a Python function and decorating it with @tool