GraphRAG: The Most Incredible RAG Strategy Revealed

[@henkhbit5748]

Great video👍 just saw today about graphrag. You're one of the first covering this. Looking forward for the next video. Graph visualization would be nice 2. Thanks.

[@KumR]

Thanks MP. Can you pl extend this to read csv, pdf, docx and add UI using streamlit too?

[@brandonvelasquez3530]

i believe it can inherently read CSV since that is basically just raw text in a specific format. I am curious about pdf and docx still

[@arifzworld]

GraphRAG is the next big thing🎉

[@TelB]

As always, Terrific - Thank you Mervin. Did I miss something about how to get the GraphRAG API? Other than that, great video. Apparently, there is a way to use this without needing the GraphRAG API. It would be great if you would cover that too please? This is not clear when you seem to state "Export the GraphRAG API Key like this". Immediately followed with, "This is the OpenAI key". I am puzzled as to whether the API Key that you use is from GraphRAG or from OpenAI please Mervin?

[@arunn2413]

What are the use cases for the text genration and embedding models? Embedding model: Indexing Text Generation:gpt-4o Summarization I think text generation is also used here for indexing, does that not involve much cost than naive RAG?

[@MrTommyorgryte]

Thanks! Great presentation as always! Can you do this using Ollama?

[@HerroEverynyan]

Thank you for the introduction so soon after the announcement! I'd be really curious to see how it compares with classic RAG on a large text where we ask for specific data, such as the taxes you'd have to pay on dividends according to the fiscal code.

[@KiwiAndCurry]

Anyone know the rough token cost for creating the relationships / user query? seems that it would likely be ~5x the cost of setting up a standard RAG.

[@oussamaoucouc6134]

You are the man , thanks again for your videos, we apreciate that

[@IdPreferNot1]

Knowledge graphs are the future... a definite component to give structure to RAG, reasoning, agentic behavior etc. That why i think LangGraph and LLamaIndex are 2 frameworks to keep up to date with.

[@awakenwithoutcoffee]

I agree but how does LangGraph relate to graphRAG?

[@dawid_dahl]

I get the feeling you didn’t exactly understand what GraphRAG is, but rather just looked at an example and made a video following that example. Didn’t get much value from this, please do a bit deeper research on a topic you are about to teach. If I wanted to read documentation, I could do that myself. My two cents. Please feel free to disagree.

[@awakenwithoutcoffee]

quite amazing isn't it. From the MS presentation it looked promising but the results took 10x more Tokens + 10x longer to generate (70s for 1 answer). How would we tackle this issue, maybe Groq inferencing could reduce the compute time ? Also: can you elaborate more on local vs global search and when to use which ? for the most accurate response maybe we should combine the two into a final answer (?). Exciting indeed, would love to see more benchmarks. 🙏

[@shawnkratos1347]

Waiting on ur next video. Please cover setting this up with ollama and openwebui

[@kylelau1329]

I can't make it work in this stage

[@panfeng2879]

Hi, Mervin, this is a very helpful video! BTW, Can I ask you a quick question? When we install the graphRAG by pip, all the necessary packages and software will be installed on our local machine, is that right? If so, what is the minimum requirements for the local machine?

[@dolife8048]

How can this be used practically inside of obsidian, where many people already have a huge database on their own fields of interest? Can you create a tutorial how to implement this in obsidian?

[@armikatollo4449]

respect bro good content. thanks

[@danishzia8192]

Hellow Mervin, when i'm intializing the grapgh rag. I'm not getting settings.yaml file. Where did you get it from?

[@kokila221]

So when you say global search, it uses the LLM in its entirety and local is just the data you uploaded... is that correct?

[@JieyiWang-xp2wh]

Why is Graph-RAG more expensive and less effective(namely, slower)? Does it have to search the whole graph for each query?

[@DEEPANMN]

Is it possible to add networkx graph into this instead of LLM generated graph! I have a readymade graph on the private dataset?

[@MrBiscuit696]

Great video! Gonna try this now

[@106rutvik]

also can you tell for what exact purpose GPT was used here? and how many tokens were you charged for?

[@john_blues]

Please when you do these , evaluate the response for correctness. That fact that it gives 'something' is not nearly sufficient.

[@gauravmodi12]

How we can see the knowledge graph on UI on Neo4j?

[@carlshod9024]

Hi! Have you try LLamaIndex Graph Rag? What are the main difference between them? Very interesting video bro

[@Jacobstalin]

Amazing content sir. This concept much much needed in current time where native RAG lacks at some point. I just wanted to ask how did you create Graph visualisation at starting? [2:57]

[@hanihaddad4317]

I have the same question too ! Did you got an answer? Thanks

[@figs3284]

Are there any ways in which you can use graphrag for coding tasks or code generation, etc? I know that wasn't their main focus with this, but I wonder if it's possible with this system.

[@106rutvik]

Hi currently we are using Pinecone Vector based DB. Can we shift to using graphrag? How it is different from vector DB? And when should we use it? Or how can we utilizes both vector DB and graph db to make outputs better?

[@60pluscrazy]

Thanks Mervin 🎉

[@ThomasConover]

5:47 Can I use this with Claude api?

[@svenandreas5947]

I wait for the ollama example .... still not sure if i got the definition of community content ..... but awesome video

[@ahmadchamseddine6891]

no ollama support

[@sharankumar31]

Hi bro kindly could you make a video on, how can i integrate this GraphRAG on phidata, crewai etc... it would be worth it...

[@rockypunk91]

If you index different documents at different point of time. We end up with multiple artifacts in the output folder. How should one do a search over all outputs. Like a production level application

[@leosong-lq4dl]

merge script

[@曾毅-y4e]

I just want to know the graphrag will extract the ner and relationship，but the original content will embed to the graphrag？hope some can reply me ❤❤

[@MidSeasonGroup]

Hi Melvin, how could this be used to optimize responses with the latest best practices and updates about a rust framework like dioxus? Many of these models are outdated and hence present a challenge.

[@ahmedbeji4248]

thank you, this is actually really exciting but is there a way to use sentence transformer embedders instead of openai or azure ? its better in my experience to use a custom embedding model trained on my data , the whole system is amazing but if its kept general it will still underperform custom systems tailored for the data If we can customise the chunking ( not token based we can actually maybe either have the chunks ready ( usually i do regex ) and use a custom transformer model ( kinda similar how u can do it in Haystack or llamaindex ) this can be really amazing

[@PriNovaFX]

Thank you, Mervin for your video and bringing this into my attention. Amazing to see that you are using Cody. What do you think, could GraphRag bring benefits to code search too?

[@MoadKISSAI]

Can we do then agentic GraphRAG? I mean having GraphRAG as a query engine tool for an agent?

[@MervinPraison]

Yes you can

[@MoadKISSAI]

@@MervinPraison are you thinking about a tutorial using multi-agents and GraphRAG? 😋

[@anandakrishnankb9172]

How can we see the graph?

[@micbab-vg2mu]

Interesting - current RAGs are not good enough for me - maybe this method will be more accurate.

[@userrjlyj5760g]

Any Local version of this, like private, without API?

[@deepanshjha6353]

First Blood 🙌

[@crippsuniverse]

Can it work with Claude 3.5 sonnet?

[@elbosanacHD]

Awesome video ! Do you know how does it compare with RAPTOR performance wise ?

[@MuhammadZubair-n7d]

Difference between local and global search is not evident through the example. I think it's assumed that the person watching the video already knows it very well.

[@shray5801]

The question here is , would this not end up in having an issue with context length?

[@dogbreath226]

Excellent intro. I've been looking forward to seeing what MS did with this research

[@theindianrover2007]

Really like your videos

[@d4rkg]

Nice video, but next time try to give a more popular source for retrieving the info, the poor gpt might probably not have any clue about such an unknown book as the one you used...

[@HassanAllaham]

This is a powerful video on a powerful tech ... waiting to see what you will do with it ...thanks for the good content 🌹🌹🌹

[@jacobriedel5326]

Does anyone know a great open source library for a chatbot that is comparable to production chatbots. A lot of enterprise level chatbots are totally lacking in the Gen AI / LLM capabilities but it would be create if developers like us could enhance a base chatbot with our own RAG techniques like GraphRAG

[@dinoscheidt]

Uhm… a chat window is simply a text field and text above it. That is so simple to do with a few lines of html that this would be a very small open source project 😅

[@MayankAnand-he8vs]

Shallow knowledge on topic

[@unclecode]

Great content! Thanks. Knowledge Graphs are superior to flat RAG systems, enabling complex queries that explore relationships between entities. They allow for more challenging questions that require connecting information, like analyzing Scrooge's actions in context. Knowledge Graphs provide structured relationships, not just text chunks, leading to more insightful answers. This approach is effective for Q&A assistants, as users seek more than just facts. Combining Knowledge Graphs with vector data is ideal. To present the real difference, instead of asking a factual question like "Who is Scrooge?", please try "what part of the story shows Scrooge doing wrong?" This requires an argument and connections between facts. Or ask, "Who is Scrooge and what is the most important thing we understand from his reaction in the story?" Such questions need to retrieve and connect information and facts.

[@agentred8732]

Your comment is as valuable as this very valuable video. A big thank you to you and to Mervin for providing such great insights into RAG and GraphRAG!

[@binaryfaith]

Can you do this demo with tabular data?

[@iham1313]

most common text form is pdf. not txt, not markdown. so how does it deal with REAL documents?

[@ULTR4_DEV]

Pdf is not a textformat

[@andrewandreas5795]

Does it work with the Claude models?

[@loicbaconnier9150]

all links reference missing 😊

[@lucface]

How does it fair with CrewAI?

[@theindianrover2007]

❤

[@adapt_it]

This content is really amazing! Thank you!

[@artur50]

anyone checked Ollama?

[@ShishirKumar07]

This is great video! Thank you Mervin.

[@deekshitht786]

Great Explanation ❤

[@evertonlimaaleixo1084]

Congrats! How much cost this process of graph generation using gpt-4o? As I understood, for each chunk you make one request to extract the relation, all right?

[@1509skate]

I just spent 38$ on a 300 page document with GPT-4o....... Wasnt even a relevant document, just a first test 😥

[@1509skate]

Just did a single Prompt against this, costet another 2.38$

[@evertonlimaaleixo1084]

@@1509skate omg!

[@artur50]

Can it be local?

[@xmagcx1]

github?

[@yuzual9506]

thx a lot for your work !

[@dhanasekar2113]

Wonderful work..

[@agentDueDiligence]

Thank you Mervin!

thanks

[@ysy69]

powerful!

[@dfwblah]

Unclear that the results are any better based on what you showed.

[@lesptitsoiseaux]

I came for the 3d graph I left empty handed.

[@alew3tube]

anybody got ollama running with graphrag?

[@alew3tube]

Ollama still doesn't support OpenAI API embeddings format, but the LLM part worked. Might need some patching to use 100% local.