RAG Explained

Get the interactive demo → ibm.biz/BdmPEb
Learn about the technology → ibm.biz/BdmPEp
Oftentimes, GAI and RAG discussions are interconnected. Learn more about about RAG is and how it works alongside your databases, LLMs and vector databases for better results with Luv Aggarwal and Shawn Brennan.
AI news moves fast. Sign up for a monthly newsletter for AI updates from IBM → ibm.biz/BdmP2c

[@amritbro]

Very simple and clear explanation.. cheers to IBM

[@paulaichniowski968]

Shawn & Luv!!!! Awesome job!!!!

[@akshitgoel1183]

Hi I have few Questions , Please find time to answer 0. Are we filtering the Data in the Vector DB ? (If Yes? then ) 1. How are we filtering the relevant data from our vector DB to augment to our prompt for the LLM ? 1. 1 who is doing this process , another LLM, our own code , or some-different tool? 1.2 Are we feeding the complete data as whole to the LLM? 1.2 if we are filtering the vector data using Rule Based Mechanism then what will be the use case of LLM , how is the power of LLM being drawn if we are the one who is making the decision what to feed as a relevant data to the LLM?

[@satish1012]

Hi This is what my understanding Storing Data as Embeddings: Correctness: Storing data (documents, images, etc.) as embeddings in a Vector DB is a valid approach. Embeddings represent the data in a high-dimensional vector space, capturing its semantic meaning. Consideration: Ensure that the embedding model you use is appropriate for your data type. For images, you might use a different model (e.g., CLIP) compared to text embeddings. Searching with Embeddings: Correctness: Converting the search query into embeddings and then comparing these embeddings with those stored in your Vector DB is correct. This allows for semantic search, which is more effective than keyword-based search. Consideration: Ensure that the conversion process and similarity calculations (e.g., cosine similarity) are implemented correctly. The returned plain text should be accurately relevant to the search query. Summarization by LLM: Correctness: Sending the retrieved plain text content to an LLM for summarization is appropriate. LLMs are designed to generate summaries and provide concise explanations based on the input text. Consideration: Ensure that the LLM is correctly configured for summarization tasks. Provide clear instructions or prompts to achieve the desired summarization quality. Returning Summarized Text to User: Correctness: Receiving the summarized text from the LLM and returning it to the user is the final step in the process. This is standard practice for providing user-friendly summaries. Consideration: Validate that the summarized content meets user expectations and provides accurate, meaningful information. 1. We Store all out relevant data in Vector DB ..like documents , images etc as part of Embeddings 2. When User Searches, it will not hit LLB directly, it will convert our search into Embeddings and return the resutlt with plain text 3. Then we send the same to text for summarization to LLM 4. Then LLM returns the summarized text back to user

[@LorenzoMarkovian]

Clear and simple, thank you guys and thank you IBM

[@emteiks]

There is good point of halucinations of AI and the video unfortunately does not address it. The data governance is not addressing this issue we still can have a scenario where input is valid but output generated by AI is a garbage.

[@vintastic_]

What's the solution?

[@frackinfamous6126]

@@vintastic_you have to make sure the relevant data is going to the model. Good info into the data base is only half the battle. Semantic chunking. Size of chunks..types of search. Type of vector database used. For example PG Vector is a Postgres plugin and is not near as good at retrieval (usually) as something like pinecone

[@frackinfamous6126]

Then the prompt used can also tremendously affect the model. You have to put it in the right context and use industry specific terms when prompting. Even a genius needs context or a bit of time to think. No matter who good the model, you have to know a bit about the specific industry to obtain great results. It’s like explains a noise to you mechanic or telling them you have a miss-fire on cylinder 1.

[@miraculixxs]

​@@vintastic_ human expert review for every response

[@aaryaxz]

Hey! If I ask a RAG-based language model, "Tell me the features of the iPhone 17," what will it tell me? Will it say it doesn't know or will it hallucinate? I understand that once the iPhone 17 is released, the database will be updated to provide the correct information. But what happens if I ask about it before its release?

[@EduardoSantAnna]

I can see two scenarios here: if it is indeed RAG-based, then you have provided info about the yet-to-be-released iPhone 17. So the LLM will respond based on that. If you don't have it in your additional documents/vector DBs, then I'd recommend you to have always added something along the lines of "only answer with facts you have access to" to your system prompt and to set Temperature to a low number. (Temperature is a parameters in LLMs that defines how "creative" the model can be) Great question and it highlights the importance of having experts in GenAI guiding enterprises on how to implement this in a way that suits their use cases.

[@fanzhang8823]

1. Vector DB may or may not stored relevant information related to the question. 2. LLM may already have information or more accurate information to the question. So RAG may not always be helpful for GenAi applications

[@vvishnuk]

Neat and detail explanation.

[@Thomas___018s]

Poetic journey: the essence of refund details and expected actions

[@mzimmerman1988]

nice work.

[@akhil7110]

Does not address how do you validate the Q1 results returned are accurate. You should have built in a process parallel to querying the LLM of actually querying the results and training the LLM to address any discrepancies, if that is possible or correct them.

[@THEaiGAI]

Awesome video

[@stephaniakossman2923]

Very clear explanation, thanks! How do you manage to avoid using "blackbox" models?

[@research2you-su9om]

Thanks for the straight forward description of RAG.

[@bayesian7404]

Good job. I still need to learn more about data accuracy in a LLM.

[@Laura__sa8q]

The demanding world of refund specifics and anticipated actions

[@CumaliTurkmenoglu-zn7hp]

Great explanation. Thank you

[@datoalavista581]

Thank for sharing !!

[@nachoeigu]

Cool explanation

[@hathanh4650]

Thank you for your sharing! Very helpful and easy to follow. Just one question, is there anyway we can test or reinforcement train the model to make sure the outputs are appropriate?

[@AgentBangla]

Excellent video, love it

[@ammularajeshsagar7953]

Informative

[@user-ht9st4up8q]

Interesting , thanks both

[@George-j3George_372u]

Ironically, because transfers between banks and cards always go so smoothly, don't they? But seriously, it's all good.

[@Patricia___361n]

Behind the scenes: Binance CEO shares insights into future developments in an exclusive interview

[@DiegoSarasua-jn2wh]

Thanks guys, very clear!

[@siddid7620]

7:10 sounds like support for open source.

[@mictow]

Did you need to learn to write backwards for this videos?? or is there a product that help you with this nice board?

[@nguyentran7068]

You record the video then mirror the image. They simple write on the clear board

[@5uryaprakashPi]

So with a rag approach, can I say that we can update the original vector db with our own processed data?

[@yinlan2672]

How to do this video? screen as the board.

[@hassanabida]

Gotcha.

[@gatsby66]

Nobody's ever been fired for buying RAGs from IBM.

[@miraculixxs]

... yet

[@ColmFearon-dc8dr]

Do these guys write on the whiteboard backwards or how does that work?

[@JShaker]

Yes they learned to write backwards for this video because it's cheaper to do that than to run the algorithm to flip a video horizontally

[@mackkaputo8989]

@@JShaker 😂😂😂

[@evgenii.panaite]

ok, gotcha 👌

[@dheerajkumar-uk6ec]

exactly love

[@topcommenter]

🤔I think Luv saw the connection the entire time

[@bangyor5949]

Do LLMs store our sensitive data when using RAG?

[@satish1012]

No , Vector DB stores all the sensitive and confidential data Once obtained send that data to LLM to get it summarized because Vector DB would have given only the connecting data for the string

[@SidStrong-l9t]

Terrible Analogy! When a journalist is ants to do research, he goes to a library and asks the librarian?? As opposed to doing a google search ? This scenario is from last century before Luv was born ?

[@BlueBearOne]

And then a wide spread global epidemic crisis is brought to light wherein our gold standard "books" (peer reviewed journals) are rife with bad and corrupt data due to mismatched incentivization and misalignment of directives; and we then realize...how much good data through science do we really have? Shame we polluted the books we are supposed to be able to trust now that we have this magnificent technology here. 😭

[@godlymajins]

Dude, keep on topic. This isn’t the place for your grievances

[@markmotarker]

lol. must be annoying to talk to "Luv". "Hi, Luv", "Exactly, Luv"

[@sk3ffingtonai]

👏👏

[@vrohan07]

kabhi haans bhi liya karo.. (Smile a bit bros)

[@maliciousinferno]

How in the world is this dude writing inverted for us too read straight lol

[@inriinriinriinriinri]

they mirror the video) you can notice that most of the people writing on a glass board are left handed(in reality 90% of planet’s population is right handed), that’s also because they mirrored the video

[@ObscuredByCIouds]

It's a skill only left-handed people have

[@Lemonsstored]

Boring