4 - Trends pt2 + Sign Language

What is prompting and what are the 2 main types?

Prompts are input instructions or queries given to a model. There is zero-shot prompting which is when you prompt without providing examples, and then there is few-shot prompting where you give a few examples to help the model understand the task.

What are some prompt design best practices?

1. Persona pattern - either the user or the LLM act as persona X to perform task Y

2. Chain-of-Though - involves making the model generate a sequence of intermediate reasoning steps before the final answer

Explain RAG

Retrieval-Augmented Generation is an apporach that combines information retrieval with language generation by supplying a language model with relevant information at inference time. Allows a system to retrieve information from external sources without retraining the model for every update

Common RAG limitations

Retrieval failure: relevant information is not retrieved

Context noise: irrelevant, outdated or redudant passages crowd the context

Common advanced RAG techniques

1. Query rewriting: Reformulates a question to improve retrieval

2. Query decomposition: Breaks a complex question into smaller questions that can be answered separately

3. Contextual compression: Removes irrelevant material from retrieved passages or condenses them

What is Contextual Engineering?

Research area that aims to ensure that a language model receives the relevant information to perform a task effectively

What are some NLP applications?

Question Answering (QA): receive a question and a context that contains information necessary to output the desired answer

Summarization: take in a document and output a summary of that document

Natural Language Inference (NLI): receive two input sentences: a premise and a hypothesis, and outputs whether the premise entails, contradicts or is neutral with respect to the hypothesis

Sentiment Analysis: models trained to classify the sentiment expressed by input text