Background Information

Project overview

Setup

Introduction

Prompt template

LCEL

Introduction to prompt engineering

Prompts

Memory

Conversation memory

Supabase

Restoring memory

Vectorstores

Supabase vector

Ingest

Retrieval

Generation

Retrieval Augmented Generation

Output Parsing

Tools

Agents

Langsmith

Farewell

( the primaryDescriptionMarkdown is long - several lines, include images, as long as you want )

In this course we <headline of what we do>. <why this technology is good> but <here's how learning it is problem>. 

In <the length of these lessons> we cover <three salient points about what we cover. Your BEST points>. Its <however long it is> and comes with <these bonus features, and reasons to buy. code, etc.>

It's taught by <author name> who <credibility, reason you should listen to them>. <social proof, if you have it yet>

<then you can list other supporting stuff you want to add, anything you want here>

-------
Example:
-------

We created a free course to create your first GraphQL server (from scratch) using Node and TypeScript. TypeScript and GraphQL are a powerful combo, but there's a lot of confusing edges if you're getting it setup the first time.

In this course we cover the basics of getting TypeScript setup with Node, how GraphQL works, and how to set up a GraphQL server using Apollo server. It's two hours long, it comes with complete code examples, and we walk through the code step-by-step.

It's taught by my friend, Shopify-engineer Hassan Djirdeh and he's integrated what he's learned using GraphQL in production as an engineer at Shopify. We've already had over 1,300 students enroll and it's been getting great reviews so far.

// ... and so on

In this course we <headline of what we do>. <why this technology is good> but <here's how learning it is problem>. 

( the summary is short - one or two lines )

Langchain.js Bootcamp

What is GPT? GPT is a language model that looks at an input token array and predicts the next token. It then appends that token and continues until a stop condition is met.

What is a token? Language models are programs and they need inputs in the format of numbers. A tokenizer turns text into number representations. Imagine that you can choose to represent the English alphabet with each character being a token from 1-26. There are different tokenizers but most of the time, tokens represent characters, words, word pieces, and special instructions.

An array of tokens is not meaningful in itself, but one of the first steps of the language model is to embed the tokens. What you need to know is that an embedding is in a consistent dimension space which means that you can calculate differences between embeddings. Similar to how we can plot things along a simple 2d graph and some points are closer together.

Tokens are also important because models typically have a maximum number of tokens they can take in called "context size". And Tokens are also what typically gets priced in APIs. APIs have gotten a lot cheaper in general, but efficiency will still matter at scale.

In addition, chat models receive additional human aided fine-tuning in a way that restricts them to performing certain tasks and prevents unwanted behavior. This is called 

Reinforcement Learning from Human Feedback (RLHF)

What goes on under the hood of an RLHF training process is complex but it's fairly easy to understand the dataset. This is a sample RLHF dataset, be warned, it's not safe for work and it may be inflammatory. Link: 

https://huggingface.co/datasets/Anthropic/hh-rlhf

Given this dataset, you can probably see how you can finetune a model yourself and there is growing support for doing so now. Both with open source models and proprietary models.

The convention for most chat models is that there is a system message, a user role, and an assistant role. These roles are used in training and are meaningful to the model.

Another model is a function model, these were trained to use helper functions like tools, and call them when provided.

More model types are likely to come out in the future so it’s best to stay up to date and connected with a community.

Let’s talk about how you can change the behavior of the model without training.

To influence the output of the model, you can only change the input, or the prompt. You can influence the prompt at each step of the process. Here are some common ways to add to the prompt:

Also one thing to note is that models often have a maximum amount of tokens accepted as input. This is called the maximum context size.

It is also useful for finding similar embeddings as we’ll see later on.

aligned models - instruct models, chat models fine tuned to do specific tasks often with special prompt formats - user, assistant, system

We will only work with aligned chat models in this course.

Most of us are familiar with databases as a persistant store of data and the way we obtain information is to query the database. There is a parallel when it comes to the internet. There is a vast amount of data in the internet, and realistically we query it through Google Search (

 of all web search is done through Google).

The data used to pre-train the language model is sourced from the internet, and so the model has learned a compressed representation of the entire internet. When you query a language model through a prompt, you are tapping into a compressed version of the internet and it is also predicting what you wish to understand about the information. Let's illustrate this.

Language models have interesting parallels with information retrieval. Most of the time, when we query in natural language we receive ranked document results. The process usually has a 

. The Query is provided in natural language. The Retrieval process is usually optimized for relevance and performance. This step usually includes something like an index.

Information Retrieval

The most ubiquitous search experience is Google Search. When you interact with google, you have to think carefully about your input. You receive a limited ranked list of results. And you have to read through the documents to satisfy your intent.

Search is actually quite similar to chatting with a model. The Query for LLMs is also natural language. And you can think of the generation step as similar to Retrieval because most of the data that Large Language Models are trained on is public data from the internet. At each instance the model is retrying it's best to complete the document instead of fetching the document.

You can also get a model to branch out and choose less likely words.

Completion

However, LLMs are not just memorizing and completing documents. This is evident because the model is not large enough to hold something like a giant switch statement. So, they have to hold some compressed representation of the training data to acheive the performance that they have.

information compression is a form of intelligence

Because LLMs are not memorizing text they do a very good job at generalizing. Although a downside to this is that language models can "Hallucinate" results that look realistic but are potentially dangerous.

The AI community is often very careful with analogies to the human mind. However, it's worth noting that Neural Networks have been around since the 60's, and were inspired by the human brain. In fact, almost everything "AI" you'll encounter in this course will probably be easy to understand based on your lived experience.

I want to caution you not to take the analogy too far. While these systems and concepts are often inspired by biology, they are also fundamentally different and are not setient.

Throughout this course, I will present a sample language application project and encourage you to think creatively. With this in mind, let's pause for a second and acknowledge that there is also great potential for wrong doing. And as developers, we have to progress responsibly.

A key idea is that LLMs have to be missing something because humans see much less data, use much less energy, and our neurons run much slower.

The pretraining process sees a large amount of internet text. An implication is that the models are trained to complete internet documents which often doesn't contain thoughtful text. There is debate on whether they can ever reach super human intelligence because they only see human text. And there is also debate on if they will ever be on par with the most capable human domain experts. There are thoughts that improving data quality or adding more data could benefit performance. For instance GPT4 has much greater reasoning than GPT3.5 and it's speculated that GPT4 was trained on a lot more code files and that improved it's performance in other domains like standardized tests. Model performance seems to grow proportionately to model parameter size and dataset size. And this relationship still hasn't hit a ceiling which suggests that bigger models and more data might be sufficient for AGI. But it's also speculated that OpenAI has run out of Github code to scrape. This brings me to the main point. There might not be enough data in the world to improve models much more. And owners are now restricting and suing AI companies that scrape text. All of this is to say that there is great potential in Prompt Engineering.

Models are very sensitive to the Prompt even after finetuning. It's not so clear why. Perhaps it triggers the completion of different kinds of documents. And it affects different models differently. There was some 

controversy surrounding Google Gemini's release

 because it outperformed GPT-4 when using Chain of Thought prompting but was less performant without prompting. However what is certain is that prompting works. It is shown in benchmarks and known in practice.

These kinds of abstract topics do not neccessarily relate to the course but if you enjoy them like these you may want to check out what researchers are thinking about.

You are likely to find out in this course that it actually takes a lot of work to get the most out of language models.

Before we move on let's talk about ethics. There are general concerns and also existential ones.

One existential concern is the runaway problem. In the runaway problem, an AI meets a threshold where it can meaningfully research AI which creates a positive feedback loop so that it becomes exponentially more intelligent. Then it has the power to do whatever it wants and potentially destroy humanity.

A general concern is model bias. The training data is skewed and that can have downstream effects on a LLM's ability to make recommendations for different demographics.

The Langchain icon, the parrot and link emoji 🦜🔗, is a reference to a 

 that warned about the dangers of biased large language models. The publishing of the paper was infamous in the AI community because it lead to the researcher, 

The parallel between chat and web search is also why Microsoft, a company that has been trying to cut into the search industry for years, became bed fellows with OpenAI and started an AI War. Geoff Hinton, who is a legend in deep learning, and the former head of Google Brain recently stepped down citing fears about AI ethics as Google has been trying to play catchup.

The pre-training process is very expensive, time consuming, and uses a lot of data. But the finetuning process is much cheaper and quality over quantity is preferred for the data. This makes finetuning much more accessible.

If you are interested in understanding the deep learning side of things, I encourage you to check out these videos afterwards. They are quite technical but I find them extremely fascinating.

The spelled-out intro to neural networks and backpropagation: building micrograd

Let's build GPT: from scratch, in code, spelled out.

[1hr Talk] Intro to Large Language Models, Andrej Karpathy

Ilya Sutskever: Deep Learning | Lex Fridman Podcast #94

David Silver: AlphaGo, AlphaZero, and Deep Reinforcement Learning | Lex Fridman Podcast #86

 is a fantastic free course that I highly recommend. It's a beginner friendly interdisciplinary course, but it stays up to date year after year so it's worth checking out again if you are familiar.

"Prerequisites assume calculus (i.e. taking derivatives) and linear algebra (i.e. matrix multiplication), we'll try to explain everything else along the way! Experience in Python is helpful but not necessary."

Theory can be helpful but it's not necessary to know everything. The goal of this course is to give you everything you need to incorporate and utilize language models into your fullstack javascript application.

There are also techniques that include adding key phrases or simply adjusting your wording. These may not always be shown on langchain, but sometimes you will encounter prompts inspired by a particular paper while browser Langchain Hub or through 

, as well as other channels. For instance, these are some key AI Researchers and influencers:

# Background Information

What is GPT? GPT is a language model that looks at an input token array and predicts the next token. It then appends that token and continues until a stop condition is met.

What is a token? Language models are programs and they need inputs in the format of numbers. A tokenizer turns text into number representations. Imagine that you can choose to represent the English alphabet with each character being a token from 1-26. There are different tokenizers but most of the time, tokens represent characters, words, word pieces, and special instructions.

An array of tokens is not meaningful in itself, but one of the first steps of the language model is to embed the tokens. What you need to know is that an embedding is in a consistent dimension space which means that you can calculate differences between embeddings. Similar to how we can plot things along a simple 2d graph and some points are closer together.

:::important
Tokens are also important because models typically have a maximum number of tokens they can take in called "context size". And Tokens are also what typically gets priced in APIs. APIs have gotten a lot cheaper in general, but efficiency will still matter at scale.
:::

In addition, chat models receive additional human aided fine-tuning in a way that restricts them to performing certain tasks and prevents unwanted behavior. This is called [Reinforcement Learning from Human Feedback (RLHF)](https://huggingface.co/blog/rlhf).

:::caution
What goes on under the hood of an RLHF training process is complex but it's fairly easy to understand the dataset. This is a sample RLHF dataset, be warned, it's not safe for work and it may be inflammatory. Link: <https://huggingface.co/datasets/Anthropic/hh-rlhf>

Given this dataset, you can probably see how you can finetune a model yourself and there is growing support for doing so now. Both with open source models and proprietary models.
:::

The convention for most chat models is that there is a system message, a user role, and an assistant role. These roles are used in training and are meaningful to the model.

Another model is a function model, these were trained to use helper functions like tools, and call them when provided.

More model types are likely to come out in the future so it’s best to stay up to date and connected with a community.

Let’s talk about how you can change the behavior of the model without training.

To influence the output of the model, you can only change the input, or the prompt. You can influence the prompt at each step of the process. Here are some common ways to add to the prompt:

-   Adding memory
-   Document retrieval
-   Giving tools to the model - (this may not be obvious, the tool descriptions are inserted into the prompt and the prompt is formatted to encourage outputting in a JSON format, which gets parsed and runs code)

:::note
Also one thing to note is that models often have a maximum amount of tokens accepted as input. This is called the maximum context size.
:::

:::tip
Playing with tokens:
<https://platform.openai.com/tokenizer>
:::

It is also useful for finding similar embeddings as we’ll see later on.

aligned models - instruct models, chat models fine tuned to do specific tasks often with special prompt formats - user, assistant, system

We will only work with aligned chat models in this course.

## Understanding Large Language Models

Most of us are familiar with databases as a persistant store of data and the way we obtain information is to query the database. There is a parallel when it comes to the internet. There is a vast amount of data in the internet, and realistically we query it through Google Search ([roughly 90%](https://www.oberlo.com/statistics/search-engine-market-share#:~:text=Top%20search%20engines%20in%202023,done%20through%20the%20internet%20giant.) of all web search is done through Google).

The data used to pre-train the language model is sourced from the internet, and so the model has learned a compressed representation of the entire internet. When you query a language model through a prompt, you are tapping into a compressed version of the internet and it is also predicting what you wish to understand about the information. Let's illustrate this.

To summarize the training process:

1.  Cover words ahead of current position.
2.  Model has to predict the next word.
3.  Model is trained to predict correct word.
4.  Move to next word and Repeat.

### Information retrieval

Language models have interesting parallels with information retrieval. Most of the time, when we query in natural language we receive ranked document results. The process usually has a **Query**, a **Retrieval**, and returns **Documents**. The Query is provided in natural language. The Retrieval process is usually optimized for relevance and performance. This step usually includes something like an index.

![Information Retrieval](https://d2uusema5elisf.cloudfront.net/courses/langchainjs-bootcamp/module_01/lesson_01.02/public/assets/InformationRetrieval.png "Information Retrieval")

:::note
The most ubiquitous search experience is Google Search. When you interact with google, you have to think carefully about your input. You receive a limited ranked list of results. And you have to read through the documents to satisfy your intent.
:::

Search is actually quite similar to chatting with a model. The Query for LLMs is also natural language. And you can think of the generation step as similar to Retrieval because most of the data that Large Language Models are trained on is public data from the internet. At each instance the model is retrying it's best to complete the document instead of fetching the document.

![LLM](https://d2uusema5elisf.cloudfront.net/courses/langchainjs-bootcamp/module_01/lesson_01.02/public/assets/LLM.png "LLM")

You can also get a model to branch out and choose less likely words.

![Completion](https://d2uusema5elisf.cloudfront.net/courses/langchainjs-bootcamp/module_01/lesson_01.02/public/assets/Completion.png "Completion")

However, LLMs are not just memorizing and completing documents. This is evident because the model is not large enough to hold something like a giant switch statement. So, they have to hold some compressed representation of the training data to acheive the performance that they have.

:::note
Some researchers believe [information compression is a form of intelligence](https://ceur-ws.org/Vol-1419/paper0045.pdf).
:::

Because LLMs are not memorizing text they do a very good job at generalizing. Although a downside to this is that language models can "Hallucinate" results that look realistic but are potentially dangerous.

## Ethics

The AI community is often very careful with analogies to the human mind. However, it's worth noting that Neural Networks have been around since the 60's, and were inspired by the human brain. In fact, almost everything "AI" you'll encounter in this course will probably be easy to understand based on your lived experience.

I want to caution you not to take the analogy too far. While these systems and concepts are often inspired by biology, they are also fundamentally different and are not setient.

Throughout this course, I will present a sample language application project and encourage you to think creatively. With this in mind, let's pause for a second and acknowledge that there is also great potential for wrong doing. And as developers, we have to progress responsibly.

### Limitations

A key idea is that LLMs have to be missing something because humans see much less data, use much less energy, and our neurons run much slower.

The pretraining process sees a large amount of internet text. An implication is that the models are trained to complete internet documents which often doesn't contain thoughtful text. There is debate on whether they can ever reach super human intelligence because they only see human text. And there is also debate on if they will ever be on par with the most capable human domain experts. There are thoughts that improving data quality or adding more data could benefit performance. For instance GPT4 has much greater reasoning than GPT3.5 and it's speculated that GPT4 was trained on a lot more code files and that improved it's performance in other domains like standardized tests. Model performance seems to grow proportionately to model parameter size and dataset size. And this relationship still hasn't hit a ceiling which suggests that bigger models and more data might be sufficient for AGI. But it's also speculated that OpenAI has run out of Github code to scrape. This brings me to the main point. There might not be enough data in the world to improve models much more. And owners are now restricting and suing AI companies that scrape text. All of this is to say that there is great potential in Prompt Engineering.

Models are very sensitive to the Prompt even after finetuning. It's not so clear why. Perhaps it triggers the completion of different kinds of documents. And it affects different models differently. There was some [controversy surrounding Google Gemini's release](https://youtu.be/90CYYfl9ntM?si=3DQgcIo0y-GSZuIS) because it outperformed GPT-4 when using Chain of Thought prompting but was less performant without prompting. However what is certain is that prompting works. It is shown in benchmarks and known in practice.

These kinds of abstract topics do not neccessarily relate to the course but if you enjoy them like these you may want to check out what researchers are thinking about.

:::info
You are likely to find out in this course that it actually takes a lot of work to get the most out of language models.
:::

### Runaway

Before we move on let's talk about ethics. There are general concerns and also existential ones.

One existential concern is the runaway problem. In the runaway problem, an AI meets a threshold where it can meaningfully research AI which creates a positive feedback loop so that it becomes exponentially more intelligent. Then it has the power to do whatever it wants and potentially destroy humanity.

A general concern is model bias. The training data is skewed and that can have downstream effects on a LLM's ability to make recommendations for different demographics.

The Langchain icon, the parrot and link emoji 🦜🔗, is a reference to a [2021 Research Paper](https://dl.acm.org/doi/10.1145/3442188.3445922) that warned about the dangers of biased large language models. The publishing of the paper was infamous in the AI community because it lead to the researcher, [Timnit Gebru](https://twitter.com/timnitGebru), being fired from Google.

:::note
The parallel between chat and web search is also why Microsoft, a company that has been trying to cut into the search industry for years, became bed fellows with OpenAI and started an AI War. Geoff Hinton, who is a legend in deep learning, and the former head of Google Brain recently stepped down citing fears about AI ethics as Google has been trying to play catchup.
:::

### Multimodality

<https://llava-vl.github.io/>

:::important
The pre-training process is very expensive, time consuming, and uses a lot of data. But the finetuning process is much cheaper and quality over quantity is preferred for the data. This makes finetuning much more accessible.
:::

## Deep Learning Resources

If you are interested in understanding the deep learning side of things, I encourage you to check out these videos afterwards. They are quite technical but I find them extremely fascinating.

## TODO <https://chat.lmsys.org/> v0-dev and redo this section

### Andrej Karpathy

[The spelled-out intro to neural networks and backpropagation: building micrograd](https://www.youtube.com/watch?v=VMj-3S1tku0)

[Let's build GPT: from scratch, in code, spelled out.](https://www.youtube.com/watch?v=kCc8FmEb1nY)

[\[1hr Talk\] Intro to Large Language Models, Andrej Karpathy](https://youtu.be/zjkBMFhNj_g?si=JzEjOD2ZJzXGvgXt)

### Lex Fridman Podcast

[Ilya Sutskever: Deep Learning | Lex Fridman Podcast #94](https://youtu.be/13CZPWmke6A?si=lqekI2uumRXxqm1k)

[David Silver: AlphaGo, AlphaZero, and Deep Reinforcement Learning | Lex Fridman Podcast #86](https://youtu.be/uPUEq8d73JI?si=EiPEvb6WaTf2d9NX)

### Academic Course

[Intro To Deep Learning from MIT OCW](http://introtodeeplearning.com/) is a fantastic free course that I highly recommend. It's a beginner friendly interdisciplinary course, but it stays up to date year after year so it's worth checking out again if you are familiar.

> **_"Prerequisites assume calculus (i.e. taking derivatives) and linear algebra (i.e. matrix multiplication), we'll try to explain everything else along the way! Experience in Python is helpful but not necessary."_** -Taken from Website

:::note
Theory can be helpful but it's not necessary to know everything. The goal of this course is to give you everything you need to incorporate and utilize language models into your fullstack javascript application.
:::

## Keeping Up with the Researchers

## TODO Keeping this section here for now

There are also techniques that include adding key phrases or simply adjusting your wording. These may not always be shown on langchain, but sometimes you will encounter prompts inspired by a particular paper while browser Langchain Hub or through [Langchain Blog](https://blog.langchain.dev/), as well as other channels. For instance, these are some key AI Researchers and influencers:

-   Andrej Karpathy, twitter: @karpathy youtube: AndrejKarpathy
-   Andrew Ng, twitter: @AndrewYNg, he also writes an AI newsletter, The Batch - <https://www.deeplearning.ai/the-batch/>
-   Aishwarya Srinivasan, twitter: @Aishwarya_Sri0, instagram: the.datascience.gal
-   Yoshua Bengio, twitter: @boredbengio
-   Yann LeCunn, twitter: @ylecun
-   Fei-Fei Li, twitter: @drfeifei
-   Lex Fridman, twitter: @lexfridman, has a wonderful podcast where he used to talk about AI a lot when it was the "AI Podcast". He still occasionally does.

    -   Wojciech Zaremba: <https://www.youtube.com/watch?v=U5OD8MjYnOM>
    -   David Silver: <https://www.youtube.com/watch?v=uPUEq8d73JI>
-   Harrison Chase
-   Jerry Liu
