I launched “Last Week in AWS” with AWS in the domain name seven years ago. AWS has never made an issue of it, though they obviously have that option.
I also have the option (and ownership) to migrate to “Last Week in the Cloud” and talk about their competitors, so it’s likely everyone is happier this way—but I confess to not kicking the bear hard enough to find out.
Most likely, the model would be less inclined to answer questions/hallucinate for prompts not related to AWS—this is definitely be a future path for improvement
Prompt: how to connect a lambda function to an RDS instance and query it safely?
1. Configure your RDS instance for outbound connections to AWS Lambda. This step involves ensuring that your DB instance is in a public subnet on your VPC and that the instance's "PubliclyAccessible" property is set to true.
Lol yeah no thanks. This is one of things that makes me nervous about using LLMs. If the internet contains shitty solutions for solving a problem, it’s going to recommend shitty solutions. Inexperienced folks won’t know better.
There’s also the thing where that’s a positive feedback look and we overflow the worlds corpus with garbage.
I've seen answers like this in Stack Overflow before. Technically correct, but basically broken in every other way: cybersecurity's future is gonna be interesting.
For that matter, using an access-key or secret-key at all is basically an anti-pattern now.
In the past few months AWS has added to their UI a popup box that asks you why you are creating access keys, it makes you answer it before generating the keys. Once you give it the answer it tells you better ways to do almost anything you're trying to do, and you have to agree to ignore those suggestions one more time before actually generating the keys.
Basically the standard access keys are really of no use anymore. You should be using roles, OIDC, IAM Identity Center, cognito federated identities, or something else. There are a million ways to identify yourself that are more secure than that and I wish tutorials stopped assuming that you should generate access keys because in no practical 2023 use case should you be doing that anymore. Especially not in any sort of sensitive or corporate environment.
If you’re part of any decently large organization, you’re probably using Control Tower or another solution that lets you create temporary credentials easily.
> If the internet contains shitty solutions for solving a problem, it’s going to recommend shitty solutions.
Have you ever noticed how the dumbest people tend be be the loudest, most confident, and most eager to share their opinions? Those are the voices that trained ChatGPT.
It’s a hard pass for me, but I fear I’m going to be forced to used it because the dumbest half of humanity will view it as a peer and use it for everything.
To connect a Lambda function to an RDS (Relational Database Service) instance and query it safely, you can follow these general steps:
1. Configure your Lambda function's execution role: Ensure that your Lambda function has the necessary permissions to access the RDS instance. The role should include the `AWSLambdaVPCAccessExecutionRole` and the appropriate RDS permissions.
2. Place your Lambda function and RDS instance in the same VPC: Ensure that both the Lambda function and the RDS instance are in the same Virtual Private Cloud (VPC) to enable their communication. You can configure the VPC settings in the AWS Management Console.
3. Configure the security group settings: Set up the security groups for your Lambda function and RDS instance to allow the necessary inbound and outbound connections. The Lambda function's security group should allow outbound access to the RDS instance's security group on the RDS port (usually port 3306 for MySQL or 5432 for PostgreSQL).
4. Create a subnet group: If your RDS instance is in a private subnet, create a subnet group that includes the appropriate subnets for your Lambda function to access the RDS instance securely.
5. Connect to the RDS instance from the Lambda function: In your Lambda function code, establish a database connection using the appropriate database client library, such as pymysql for MySQL or psycopg2 for PostgreSQL. Use the connection details (endpoint, username, password, etc.) of your RDS instance.
6. Execute queries safely: To query the RDS instance safely, ensure that your Lambda function's code incorporates secure coding practices. Use parameterized queries or prepared statements to prevent SQL injection attacks. Avoid storing sensitive information (such as database credentials) in your Lambda function code and instead use environment variables or AWS Secrets Manager for secure storage and retrieval.
Remember to regularly update and patch your Lambda function, RDS instance, and associated dependencies to maintain security and stay protected against potential vulnerabilities.
It's worth noting that the exact implementation details may vary depending on your specific use case, programming language, and AWS setup. Consulting the AWS documentation and resources related to Lambda, RDS, and VPC networking can provide more detailed and up-to-date instructions for your specific requirements.
It still isn't tho, this will work for testing, maybe, but each lambda will be it's own connection and that will exhaust db resources real fast, you're supposed to have a pooling proxy between lambdas and RDS.
I asked ChatGPT the same question again and then followed up with:
“Is there a more efficient way to connect to the database”
It then went off the rails. It gave me generic answers about using connection pooling that your language framework provides.
Of course that doesn’t help with each Lambda being invoked separately.
I then asked “But each Lambda runs based on its own vm”
It then correctly said that while each Lambda invocation would have its own pool and couldn’t be shared, it could be shared with multiple connections during one invocation.
Which is technically correct. But not the answer I was looking for.
In all fairness, it’s the same type of answers I would expect from a junior to mid developer with limited experience with AWS. I would hire a person who could answer that well.
I try not to cheat and hint at specific issues (since that relies on prior knowledge)
I'd be surprised if even after the last prompt it wouldn't notice that. Saying "Did we miss anything" leaves it open it to re-evaluate both the implementation and the original considerations
This to me is why I don't use ChatGPT. Every time I use it I get answers that are questionable in validity. I'd say that 3/4 of my questions have returned incorrect answers at least partially. Maybe 10% are wildly incorrect with the remainder being partially incorrect.
So if it is wrong with MOST of the questions that I am able to validate myself, then how can I trust it on the questions that I am unable to validate myself.
The whole reason I started even doing this passive research is because I had an employee late last year who used to be a decent employee (he was never an all-star, but he got the job done to an average/satisfactory level), and started to all of a sudden perform incredibly poorly.
He was submitting code and solving problems that were just really bad. He was always just an average producer, and was always teetering on the edge during performance reviews, doing just well enough that we kept him around. But he quickly started to make mistake after mistake. Several code reviews I found really strange artifacts and comments in his code that were blatant mistakes. I confronted him about them over the course of several performance reviews and said he blamed stackoverflow "copy pasta". Eventually I actually fired him, since it was getting to a point that almost everything he submitted or produced was problematic in some way and he was burning more of my time than he was saving. So I ended up firing him.
While firing him, in front of HR he finally broke down and admitted that he has been using ChatGPT for everything and he begged us to let him stay and he would stop using it altogether. I of course didn't care at that point and we let him go. But I started to realize the increase of mistakes were all due to ChatGPT leading him astray.
That whole experiment really taught me that ChatGPT is not ready for primetime. If you blindly trusted ChatGPT you will find yourself in the wrong place most of the time. The problem is that unless you already know the answer to the question you are asking, it is very difficult to tell where chatgpt's answers might be correct and were they might be incorrect (because it is usually a mixture of both). This makes it entirely useless for asking questions that you are not comfortable validating.
Why’d he be so reluctant to confess to using ChatGPT? And why would he continue using it despite constant feedback that his performance was getting so much worse?
> So if it is wrong with MOST of the questions that I am able to validate myself, then how can I trust it on the questions that I am unable to validate myself.
If you’re using it to generate code, you can validate it yourself - run the code.
Presumably you're a good manager but the way you told it makes you sound like a bad manager that didn't dig into what's going on with a poor performing employee and let them flail around until you had to fire them. There's more to that story and it's not really about ChatGPT at all.
There are other stories out there, like https://hyperbo.la/w/chatgpt-4000/ which shows it can useful and a force multiplier when used well, but it's like giving a faster car to a bad driver. It'll just result in them crashing faster. If you've got a programmer that doesn't want to program, ChatGPT can't help them be a better programmer since they don't actually want to be one!
ChatGPT doesn't need to provide perfect AWS solutions for people without AWS experience; the onus is on the human to do the work necessary on their side to be sure of the solution it produces.
It should be thought of as a time saving tool for experts (and people willing to put the work in), not a magic button for lazy people.
The role doesn’t need any IAM RDS permissions. It needs network access. But then all of the access to the database is controlled by your standard database permissions - not IAM.
I guess there are corner cases like using IAM permissions to connect to your database which can be done. But is not typical
They're all so eager to please they will basically never say "that's actually not possible", and invent some plausibly sounding bullshit.
For beginners who are struggling to tell the difference between 1/ how to do something that's possible, 2/ how to do something that SHOULD be possible but just currently isn't, and 3/ how to do something that is RIDICULOUS and shouldn't even ever be possible, ChatGPT is worse than nothing :(
It seems to know more than just AWS, I was thinking it was just embeddings but then I asked it “Is Aws better than azure?” And it seemed to give an answer which widener seem to be derived from the source documents.
The thought that a bunch of people will trust this tool and make some terrible decisions is unsettling. On the other hand it could be pretty powerful if you know what you are doing.
Using ChatGPT for AWS service questions is actually pretty good. For instance, I asked it for a Cloud Practitioner study guide (using a small set of crafted prompts), and GPT performed quite well. While I have yet to query GPT about Solutions Architect or DevOps material, I know I can feed a set of URLs and GPT will "learn" the material and summarize it in ways meaningful and relevant to my prompts. In this way, ChatGPT is quite a powerful assistant on its own.
Agree, I've had great results asking chatgpt questions about AWS services. The interactivity is very helpful, and chatgpt will draft scripts too, although that's hit or miss. But for understanding concepts and services, it's great.
Meh. It doesn't actually pull a valid response. We just upgraded a database that required an updated EC2 instance, and it failed to connect those dots.
This is exactly the kind of software that people should be making. Even if this one is bad, the concept is very sound.
I want to have a specialist AI that is trained to help me learn how to use the software. 100% what should be happening.
General AI should know how to do stuff too but having an specialist AI implies that the company/group of people/person is making sure to tune the model.
Here's one example question that ChatGPT utterly fails at, but that this answers fine: "What is Amazon CodeCatalyst?"
ChatGPT: "I'm sorry, but as of my knowledge cut-off in September 2021, there was no service, tool, or product known as Amazon CodeCatalyst offered by Amazon Web Services (AWS). [...]"
ChatGPT is not up to date unless you start using the plugins. This sort of indexing is based on vector databases and various intermediate prompting. If you want to get technical, the academic term is "Retrieval Augmented Generation".
Hallucination is unfortunately inevitable when it comes to any autoregressive model, even with RAG. You can minimize hallucination by prompting, but you'll still see some factually incorrect responses here and there (https://zilliz.com/blog/ChatGPT-VectorDB-Prompt-as-code).
I unfortunately don't think we'll be able to solve hallucination anytime soon. Maybe with the successor to the transformer architecture?
Hallucination is naturally a concern for anyone looking to depend upon LLM-generated answers.
We’ve been testing LLM responses with a CLI, we’re using it to generate accuracy statistics, which is especially useful when the use-case Q/A is limited.
If ‘confidence’ can be returned to the user, then at least they can have an indication if there is a higher quality-risk with a given response.
Is it? It's true there are a lot of very loud messages on that linked issue, but only MDN knows how many people are using and engaging with their system compared to on that thread. We, as the general public, don't know anything other than there's a vocal group of dissidents, and what MDN tells us. Which I looked and didn't find stats on its usage.
Hey everyone, I am the creator of AWS Docs GPT and its been extremely useful to garner all of your feedback for the site—thank you guys so much! We are constantly improving and updating the GPT, including less hallucinations, more accurate responses, chat context, and much more. At the end of the day, I really hope this tool can be useful for developers like myself out there!!!
What's the best current way to make a bunch of documents searchable via LLMs like this?
I've tried the using OpenAI w embeddings (iirc), but this was slow, got expensive quickly, and it struggled to answer questions about the text accurately. Curious if there's better standard approaches now.
A couple of things come to mind:
1. embedding methods: there're a couple of ways to do that, the most used one is OpenAi's text-davinci-002, although in my use case(short sentence description of API) it didn't work pretty well;
2. how you truncate documents into pieces: for this langchain has some implementation and helpful pointers.
I think you have to do lots of experiment on this till you find your best information retrieval strategy
I really love this concept. While I do get better results from GPT-4 for AWS questions right now, AI as the "interpreter" for documentation works really well.
I've been using GPT4 for this since the beginning - ironically a large majority of AWS documentation has been machine generated since 2018. Circa 2019 the entire exlixir API for AWS was machine generated.
Asking GPT4 is also consistently less of a headache than asking the devops guy and getting a 20min explanation for a simple question.
I am building something similar, it has documentation from azure, aws, and lots of slack/discord threads of software projects are also searchable, check it out https://www.kwq.ai
it even gets real time indexing from slack of aws deep java library, and from discord of deepset haystack project
No, AWS is not designed to take all your money. AWS offers a variety of payment options and cost-saving measures to help you manage your expenses effectively. ...
RAG is very difficult to do right. I am experimenting with various RAG projects from [1]. The main problems are:
- Chunking can interfer with context boundaries
- Content vectors can differ vastly from question vectors, for this you have to use hypothetical embeddings (they generate artificial questions and store them)
- Instead of saving just one embedding per text-chuck you should store various (text chunk, hypothetical embedding questions, meta data)
- RAG will miserably fail with requests like "summarize the whole document"
- to my knowledge, openAI embeddings aren't performing well, use a embedding that is optimized for question answering or information retrieval and supports multi language. SOTA textual embedding models can be found on the MTEB Leaderboard [2]. Also look into instructorEmbeddings
- the LLM used for the Q&A using your context should be fine-tuned for this task. There are several open (source?) LLMs based on openllama and others, that are fine tuned for information retrieval. They hallucinate less and are sticking to the context given.
I’ve been working with RAG for months, too, and it’s vanishingly rare to see anything but toy examples in the wild. This is a solid, concise list of where the dragons are.
Any idea where all the RAG practitioners hang out and trade war stories? Is there a forum or Discord or something?
To solve the question embedding issue I prefer another approach, you store document embeddings as normal, then for searching you let LLM hallucinate an answer and use the wrongish answer embedding to do the vector search.
The inverse idea of Hypothetical Embeddings, HyDE [1]
"HyDE is an embedding technique that takes queries, generates a hypothetical answer, and then embeds that generated document and uses that as the final example."
BriefGPT [2] is implementing this and it uses the following prompt at ingestion-time:
"Given the user's question, please generate a response that mimics the exact format in which the relevant information would appear within a document, even if the information does not exist.
The response should not offer explanations, context, or commentary, but should emulate the precise structure in which the answer would be found in a hypothetical document.
Factuality is not important, the priority is the hypothetical structure of the excerpt. Use made-up facts to emulate the structure.
For example, if the user question is "who are the authors?", the response should be something like
'Authors: John Smith, Jane Doe, and Bob Jones'
The user's question is:"
Marqo provides automatic, configurable chunking (for example with overlap) and can allow you to bring your own model or choose from a wide range of opensource models. I think e5-large would be a good one to try. https://github.com/marqo-ai/marqo
Using Qdrant doesn’t require docker, like Marqo does (from the README). Any trade offs between the two? Doc chunking is an independent functionality and there are already line that help chunk with overlap etc,and also it’s not hard to roll your own.
That’s an interesting list (awesome-marketing-data science). Curious what is specific to marketing in that list, or maybe it’s just an inaccurate name.
historical repo name, it's really not that anymore, besides a very old list of marketing stuff that i rarely update. I should rename the repo, but I hesitate :)
133 comments
[ 2.3 ms ] story [ 204 ms ] threadI originally thought this was an official Amazon website...their lawyers would probably say the same thing...
I launched “Last Week in AWS” with AWS in the domain name seven years ago. AWS has never made an issue of it, though they obviously have that option.
I also have the option (and ownership) to migrate to “Last Week in the Cloud” and talk about their competitors, so it’s likely everyone is happier this way—but I confess to not kicking the bear hard enough to find out.
I’m sure you know that your name is brought up frequently inside AWS.
how do you use wrangler and glue to make athena tables using terraform
Glacier is a term that is not directly mentioned in the provided sources.
Prompt: What is a glacier?
A glacier is a large mass of ice that moves slowly over time due to the accumulation of snow, ice, and other forms of frozen precipitation.
Seems like it’s just using a general model?
When the docs are wrong or misleading you'll still get burned, even if the model doesn't hallucinate responses
Prompt: how to connect a lambda function to an RDS instance and query it safely?
1. Configure your RDS instance for outbound connections to AWS Lambda. This step involves ensuring that your DB instance is in a public subnet on your VPC and that the instance's "PubliclyAccessible" property is set to true.
There’s also the thing where that’s a positive feedback look and we overflow the worlds corpus with garbage.
I've seen answers like this in Stack Overflow before. Technically correct, but basically broken in every other way: cybersecurity's future is gonna be interesting.
In the past few months AWS has added to their UI a popup box that asks you why you are creating access keys, it makes you answer it before generating the keys. Once you give it the answer it tells you better ways to do almost anything you're trying to do, and you have to agree to ignore those suggestions one more time before actually generating the keys.
Basically the standard access keys are really of no use anymore. You should be using roles, OIDC, IAM Identity Center, cognito federated identities, or something else. There are a million ways to identify yourself that are more secure than that and I wish tutorials stopped assuming that you should generate access keys because in no practical 2023 use case should you be doing that anymore. Especially not in any sort of sensitive or corporate environment.
Have you ever noticed how the dumbest people tend be be the loudest, most confident, and most eager to share their opinions? Those are the voices that trained ChatGPT.
It’s a hard pass for me, but I fear I’m going to be forced to used it because the dumbest half of humanity will view it as a peer and use it for everything.
To connect a Lambda function to an RDS (Relational Database Service) instance and query it safely, you can follow these general steps:
1. Configure your Lambda function's execution role: Ensure that your Lambda function has the necessary permissions to access the RDS instance. The role should include the `AWSLambdaVPCAccessExecutionRole` and the appropriate RDS permissions.
2. Place your Lambda function and RDS instance in the same VPC: Ensure that both the Lambda function and the RDS instance are in the same Virtual Private Cloud (VPC) to enable their communication. You can configure the VPC settings in the AWS Management Console.
3. Configure the security group settings: Set up the security groups for your Lambda function and RDS instance to allow the necessary inbound and outbound connections. The Lambda function's security group should allow outbound access to the RDS instance's security group on the RDS port (usually port 3306 for MySQL or 5432 for PostgreSQL).
4. Create a subnet group: If your RDS instance is in a private subnet, create a subnet group that includes the appropriate subnets for your Lambda function to access the RDS instance securely.
5. Connect to the RDS instance from the Lambda function: In your Lambda function code, establish a database connection using the appropriate database client library, such as pymysql for MySQL or psycopg2 for PostgreSQL. Use the connection details (endpoint, username, password, etc.) of your RDS instance.
6. Execute queries safely: To query the RDS instance safely, ensure that your Lambda function's code incorporates secure coding practices. Use parameterized queries or prepared statements to prevent SQL injection attacks. Avoid storing sensitive information (such as database credentials) in your Lambda function code and instead use environment variables or AWS Secrets Manager for secure storage and retrieval.
Remember to regularly update and patch your Lambda function, RDS instance, and associated dependencies to maintain security and stay protected against potential vulnerabilities.
It's worth noting that the exact implementation details may vary depending on your specific use case, programming language, and AWS setup. Consulting the AWS documentation and resources related to Lambda, RDS, and VPC networking can provide more detailed and up-to-date instructions for your specific requirements.
“I’m a beginner. Walk me through step $n”
“Is there a more efficient way to connect to the database”
It then went off the rails. It gave me generic answers about using connection pooling that your language framework provides.
Of course that doesn’t help with each Lambda being invoked separately.
I then asked “But each Lambda runs based on its own vm”
It then correctly said that while each Lambda invocation would have its own pool and couldn’t be shared, it could be shared with multiple connections during one invocation.
Which is technically correct. But not the answer I was looking for.
In all fairness, it’s the same type of answers I would expect from a junior to mid developer with limited experience with AWS. I would hire a person who could answer that well.
- What are the considerations?
- Implement it with the considerations
- Did we forget anything?
It understands the need for a proxy from step 1:
https://chat.openai.com/share/7ca37130-a771-457c-8742-a5f941...
It still missed using a paginator to handle the list_roles call returning more than 50 roles.
Once I pointed it out, it did add pagination support.
I'd be surprised if even after the last prompt it wouldn't notice that. Saying "Did we miss anything" leaves it open it to re-evaluate both the implementation and the original considerations
Edit: There's some non-determinism involved, but GPT-4 caught the pagination from planning stage here: https://chat.openai.com/share/3c356d4f-15d4-4f6e-bd29-af6a0b...
https://aws.amazon.com/about-aws/whats-new/2020/06/amazon-rd...
So if it is wrong with MOST of the questions that I am able to validate myself, then how can I trust it on the questions that I am unable to validate myself.
The whole reason I started even doing this passive research is because I had an employee late last year who used to be a decent employee (he was never an all-star, but he got the job done to an average/satisfactory level), and started to all of a sudden perform incredibly poorly.
He was submitting code and solving problems that were just really bad. He was always just an average producer, and was always teetering on the edge during performance reviews, doing just well enough that we kept him around. But he quickly started to make mistake after mistake. Several code reviews I found really strange artifacts and comments in his code that were blatant mistakes. I confronted him about them over the course of several performance reviews and said he blamed stackoverflow "copy pasta". Eventually I actually fired him, since it was getting to a point that almost everything he submitted or produced was problematic in some way and he was burning more of my time than he was saving. So I ended up firing him.
While firing him, in front of HR he finally broke down and admitted that he has been using ChatGPT for everything and he begged us to let him stay and he would stop using it altogether. I of course didn't care at that point and we let him go. But I started to realize the increase of mistakes were all due to ChatGPT leading him astray.
That whole experiment really taught me that ChatGPT is not ready for primetime. If you blindly trusted ChatGPT you will find yourself in the wrong place most of the time. The problem is that unless you already know the answer to the question you are asking, it is very difficult to tell where chatgpt's answers might be correct and were they might be incorrect (because it is usually a mixture of both). This makes it entirely useless for asking questions that you are not comfortable validating.
If you’re using it to generate code, you can validate it yourself - run the code.
There are other stories out there, like https://hyperbo.la/w/chatgpt-4000/ which shows it can useful and a force multiplier when used well, but it's like giving a faster car to a bad driver. It'll just result in them crashing faster. If you've got a programmer that doesn't want to program, ChatGPT can't help them be a better programmer since they don't actually want to be one!
It should be thought of as a time saving tool for experts (and people willing to put the work in), not a magic button for lazy people.
> and the appropriate RDS permissions.
The role doesn’t need any IAM RDS permissions. It needs network access. But then all of the access to the database is controlled by your standard database permissions - not IAM.
I guess there are corner cases like using IAM permissions to connect to your database which can be done. But is not typical
https://imgur.com/a/0IYZ2WV
https://www.phind.com/search?cache=d0b3a85b-17f9-4def-b8d0-b...
They're all so eager to please they will basically never say "that's actually not possible", and invent some plausibly sounding bullshit.
For beginners who are struggling to tell the difference between 1/ how to do something that's possible, 2/ how to do something that SHOULD be possible but just currently isn't, and 3/ how to do something that is RIDICULOUS and shouldn't even ever be possible, ChatGPT is worse than nothing :(
I want to have a specialist AI that is trained to help me learn how to use the software. 100% what should be happening.
General AI should know how to do stuff too but having an specialist AI implies that the company/group of people/person is making sure to tune the model.
Just an IMO.
You can ask the same questions to ChatGPT and get the same or better answers.
I also know from personal experience with ChatGPT, that you can use it to:
- convert Python/boto3 to any language that has an AWS SDK
- convert CloudFormation to Terraform or the CDK
- write scripts that use the SDK
You will get the occasional hallucination.
ChatGPT: "I'm sorry, but as of my knowledge cut-off in September 2021, there was no service, tool, or product known as Amazon CodeCatalyst offered by Amazon Web Services (AWS). [...]"
I unfortunately don't think we'll be able to solve hallucination anytime soon. Maybe with the successor to the transformer architecture?
We’ve been testing LLM responses with a CLI, we’re using it to generate accuracy statistics, which is especially useful when the use-case Q/A is limited.
If ‘confidence’ can be returned to the user, then at least they can have an indication if there is a higher quality-risk with a given response.
[1] https://docs.cloudlinux.com
https://github.com/mdn/yari/issues/9208
https://imgur.com/a/IGu1syf
I've tried the using OpenAI w embeddings (iirc), but this was slow, got expensive quickly, and it struggled to answer questions about the text accurately. Curious if there's better standard approaches now.
https://github.com/Mintplex-Labs/anything-llm
I think you have to do lots of experiment on this till you find your best information retrieval strategy
Asking GPT4 is also consistently less of a headache than asking the devops guy and getting a 20min explanation for a simple question.
These are fun projects!
How does this work?
it even gets real time indexing from slack of aws deep java library, and from discord of deepset haystack project
No, AWS is not designed to take all your money. AWS offers a variety of payment options and cost-saving measures to help you manage your expenses effectively. ...
- Chunking can interfer with context boundaries
- Content vectors can differ vastly from question vectors, for this you have to use hypothetical embeddings (they generate artificial questions and store them)
- Instead of saving just one embedding per text-chuck you should store various (text chunk, hypothetical embedding questions, meta data)
- RAG will miserably fail with requests like "summarize the whole document"
- to my knowledge, openAI embeddings aren't performing well, use a embedding that is optimized for question answering or information retrieval and supports multi language. SOTA textual embedding models can be found on the MTEB Leaderboard [2]. Also look into instructorEmbeddings
- the LLM used for the Q&A using your context should be fine-tuned for this task. There are several open (source?) LLMs based on openllama and others, that are fine tuned for information retrieval. They hallucinate less and are sticking to the context given.
1 https://github.com/underlines/awesome-marketing-datascience/...
2 https://github.com/embeddings-benchmark/mteb
I’ve been working with RAG for months, too, and it’s vanishingly rare to see anything but toy examples in the wild. This is a solid, concise list of where the dragons are.
Any idea where all the RAG practitioners hang out and trade war stories? Is there a forum or Discord or something?
BriefGPT [2] is implementing this and it uses the following prompt at ingestion-time:
"Given the user's question, please generate a response that mimics the exact format in which the relevant information would appear within a document, even if the information does not exist. The response should not offer explanations, context, or commentary, but should emulate the precise structure in which the answer would be found in a hypothetical document. Factuality is not important, the priority is the hypothetical structure of the excerpt. Use made-up facts to emulate the structure. For example, if the user question is "who are the authors?", the response should be something like 'Authors: John Smith, Jane Doe, and Bob Jones' The user's question is:"
1 https://python.langchain.com/docs/modules/chains/additional/...
2 https://github.com/e-johnstonn/BriefGPT