ChatGpt Series 1: Model Introduction

GPT4\GPT-3.5

The multimodal model is a natural language processing model that can understand text and is very human-like. It is not a common instruction-type fake model. It is a human-like model that mainly receives text input and outputs text.

6 Strategies for Better Results

Write clear instructions

Strategy:

• Require the model to adopt a role (system messages can be used to specify the role that the model uses in its responses. This is necessary when we implement the application. Set the role to our own project name)
• Use delimiters to clearly indicate different parts of the input
• Specify the steps to complete the task
• Provide examples (for example, when doing intelligent customer service, try to have a unified response style every time you answer, and you can provide example implementations)
• Specify the desired output length

Provide reference text

Strategy

• Instructs the model to answer using reference text

• Instructs the model to answer by citing the reference text

Split complex tasks into simpler subtasks

Strategy

• Use intent classification to identify instructions most relevant to user queries
• For conversation apps that require long conversations, summarize or filter previous conversations
• Summarize long documents segmentally and recursively build a full summary

Give GPT time to think

Strategy

• Instruct the model to find its own solution before rushing to conclusions

• Use internal monologue or a series of queries to hide the model’s reasoning process

• Ask the model if it missed anything in the previous pass

Use external tools

Compensate for GPT's weaknesses by feeding it the output of other tools

Strategy

• Use Embeddings-based search to achieve efficient knowledge retrieval (functions based on local data search can be implemented using Embeddings, such as quickly building intelligent help centers, customer service systems, consultation systems, insurance recommendations, etc.)

• Use code execution to perform more accurate calculations or call external APIs

  For example: you can provide external calling interfaces or logic codes in the gpt context to quickly get answers. This is very effective for integrating your own business.

• Grant a model access to specific functionality

System testing changes

There needs to be a way to test whether an optimization is a positive improvement to the system. A good evaluation method is:

• Represents real-world usage
• Specify test cases
• Easy to automate or repeat

Strategy: Evaluate model output against gold standard answers

It is assumed that the correct answer to a known question should refer to a specific set of known facts. We can then use model queries to calculate how many required facts are included in the answer.

*The above strategies are just tips, you can knead and use them as needed

FROM E/DALL.E2

Create images from natural language descriptions. The most direct landing scenario:

• Promotional poster generation system (removing the UI is equivalent to configuring a 1v1 UI designer for each salesperson to improve the salesperson conversion rate)

The Images API provides three methods for interacting with images:

1. Create images from scratch based on text prompts
2. Edit existing images with new text prompts
3. Create a variation of an existing image

Whisper

Speech recognition model can be applied to:

• Transcribe the audio into any language the audio is in;
• Translate and transcribe audio into English;

The current limit for text items is 25 MB, and the following import text type: mp3, mp4, mpeg, mpga, m4a, wavsumwebm.

Common application scenarios:

• Automatically generate meeting minutes (online meeting application scenario)
• Text novels converted into audio readers (Himalayan type)
• Douyin mass-produces storytelling-type materials

And we can define our company and product names by specifying system_prompt of GPT4

Embeddings

Embedding is usually used in the following scenarios:

• Search(results sorted by query string relevance)
• Clustering(Group text strings by similarity)
• Recommend(Recommend items with related text strings)
• Anomaly detection(Identify outliers with low correlation)
• Diversity measurement(Analysis of similarity distribution)
• Classification(Text strings are classified by their most similar tags)

To put it simply, Embeddings can convert the provided data into vector data and then apply it to the above scenarios.

The most intuitive application scenario is smart customer service and smart recommendation. It should be a scenario that is relatively easy to implement and can achieve immediate results. Combined with the qdrant database (or Tiger Balm Redis, too), local data can be converted through Embeddings and stored in qdrant to implement local search.

Or make user portraits, and after converting user data through Embeddings, you can easily automatically tag users. An artificial intelligence version of the user tagging system is completed. Then empower it to other systems (CRM, recommendation system, etc.).

In the same way, intelligent recommendation systems can also be implemented very quickly through Embeddings. Just adjust the parameters or add some additional code logic, and an intelligent system that conforms to the company's business is completed.

Vector database options include:

• Chroma, an open source embed store
• Milvus, a vector database built for scalable similarity searches
• Pinecone, a fully managed vector database
• Qdrant, vector search engine
• Redis serves as a vector database (small storage capacity, generally not used)
• Typesense, fast open source vector search
• Weaviate, an open source vector search engine
• Zilliz, data infrastructure powered by Milvus

Moderation (free)

This is the one we are most familiar with, the review model. From now on, pinyin abbreviations like yyds will no longer be able to escape. It’s really yyds.