KAG Framework SMASHES GraphRAG in Accurate Knowledge Generation

[@annonymbruger]

You drop this on the 24’ - my Christmas Eve is spoiled 😂 I need to test this out now :)

[@alwikah5664]

Don't forget to start with "This is amaaaazing" 🙂 Top video, thanks for sharing.

[@joseduarte1240]

What do you think its the best way if we want to create an TEXT to SQL with a complex database?

[@sergeziehi4816]

Thanks Mervin; Would be good if it retrieves also references along with question answering in the retrival process.

[@patruff]

The key differentiator is that KAG puts more emphasis on knowledge representation and reasoning capabilities, while LightRAG focuses more on efficient retrieval through its dual-level approach. For applications requiring deep knowledge understanding and complex reasoning, KAG would likely be the better choice. For applications prioritizing fast retrieval and simpler queries, LightRAG may be sufficient. I recommend KAG for use cases requiring: - Complex multi-hop reasoning - Deep domain knowledge integration - Sophisticated knowledge representation needs And LightRAG for use cases requiring: - Fast retrieval performance - Simpler knowledge representation needs - Efficient processing of straightforward queries

[@ysy69]

When you say fast retrieval, how fast is it compared to KAG retrieval?

[@RobertFitz-r8r]

Thanks, I downloaded KAG and played around with it after I saw your video. Unfortunately, the KAG document processing fails every time during the text embedding phase (step 4) with the error message - cannot connect to OpenAI. I configured it to use the OpenAI text embedding model. KAG managed to connect to the OpenAI LLM (in my case gpt-4o-mini), that worked. But it failed to invoke the OpenAI embedding model. It still tried its best to answer my query in the chat window, but I could tell from the response that the additional document embedding was missing. It’s unfortunate that the product quality is in early beta phase. I would have really liked to experiment more with it.

[@alprbgt]

Hi bro, did you solve that problem?

[@RobertFitz-r8r]

@@alprbgt No, I am waiting until the authors are updating their repo to address the issue. There are many other interesting solutions, like LightRAG, that do work - so no need to waste time to investigate a solution that currently does not work.

[@ysy69]

Grateful for informing and tutoring us. Does this mean we can leave RAG/GraphRAG behind and move full steam ahead with KAG?

[@SullyOrchestration]

You guys did fantastic work. Is there a way for this to be deployed as a self-hosted API package? For example is it possible if I create an AI agent with n8n and connect it to this? This would be the best tutorial EVER!

[@takimdigital3421]

Still no community node , but you could run python script and retrieve the answer

[@ahmadzaimhilmi]

My issue with KG RAG is that I'm unable to reference any snippet from the document the result is retrieved. The result may be accurate, but what's the point of using it if we need to go back and document between our document and the generated result to verify the facts?

[@SullyOrchestration]

Agreed!

[@pink_fluffy_sky]

can we combine this with traditional rag?

[@ysy69]

What do you mean you're unable to reference any snippet from the document?

[@SuperLano98]

I still don’t understand what’s the difference and advantages of KAG over graph rag. Could you, please, clarify it ?

[@janalgos]

How does this compare to hybrid rag approach? The one that mixes graph rag with vector rag

[@01of13]

thanks for the video. would you mind creating a version that shows how to configure KAG to work with openai models? it is hard to copy the config code from the video, as it is not provided separately as you customarily do.

[@alprbgt]

Hi, First of all, Thanks for your good video. It's knowledgeble video. I tried what you do in the video on my computer, but when I create a task, I always getting a vectorization connection error. I'm using OpenAI API and it's embedding model. My question, is there any document or any video on platforms about that error?

[@shaunlee8529]

im not sure why, but the image i used yours for spg server, but im still required to log in with credentials. its a local setup, so im not sure where can i get the credentials from

[@crimsonkim6824]

Do you know if it even out-performs Lazy GraphRAG from Microsoft? Cuz I thought they were currently considered kinda like the SOTA rn in the Graph RAG paradigm

[@AbdulSamada-qv7se]

when i run the url its asking for login >\?

[@belfort777]

@mervin please make KAG with DeepSeek-R1

[@Kevencebazile]

Just Reviewed amazing content happy holidays !! thanks for content

[@SullyOrchestration]

I am trying this now - the knowledge extraction process seems to be extremely slow with a chunk size of 4060 and length of 500 each. In fact one chunk had to be processed for 10 minutes before moving on to the next chunk. While I appreciate that I have a large document, it would be nice to see more detailed logs or estimated time to completion

[@0730pleomax]

That slow?

[@milindtakate5987]

Can we do the incremental update to the graphs created?

[@zhengkegui]

Yes, kag uses neo4j as the graph storage engine, and each graph construction task writes incremental data to the knowledge base.

[@JorgeCastro-vm1kt]

Thanks Mervin!

[@davidjourno6929]

Hey Mervin, I'm currently learning KAG, I'm assuming KAG can be used with sparql but for indexing, I know this is pre built framework but I'm curious how performs something different like elasticsearch vectors or vec2sparql

[@0x7f6]

Nice explanation and demo, Mervin. Can the queries work across any/all of the uploaded documents rather than one specific document?

[@dekel56]

anyone knows how to write the model configuration for openai? I get connection error 2024-12-24 23:23:42: execute error:java.util.concurrent.ExecutionException: pemja.core.PythonException: : Connection error.

[@BryanAbrilSar]

How I can use this framework with n8n it can ?

[@mohamedfouad1309]

This is amazing

[@scitechtalktv9742]

Fantastic! I just discovered this. May I ask exactly how you did this? I mean technically, because I know that these 2 songs are compatibel according to music theory (Camelot Wheel)

[@adanpalma4026]

Thanks, so wich one is the way to go? Agentic Rag(langgraph, phidata), llamaparse and semantic? Or this one. I am really confuse how to go. I just want to build good rad, eficient, with few o no hallucination but I am really lost

[@zhengkegui]

If there were end-to-end quantitative evaluations in specific fields, it would be easier to choose a technical route. If you know of such methods, please let me know. According to KAG Technical report, it seems to perform well on multi-hop question answering datasets like HotpotQA, TwoWiki, and MuSiQue.

[@adanpalma4026]

The first thing is how good is parsing. LlamaIndex offer llamaparse that claim to be the best one or at least one of the best. I want to combina llamaparsw+chromadb or+ agentic rag (phidata or langgraph) or kag but for me is overwhelming

[@adanpalma4026]

@@zhengkegui I have bee practicing with llamaindex+llamaparse+chromadb. Also with phidata with agents and first steps with langgraph with agents, I have to admit are small steps and No in a serious or businness solution, but I am really confused becasuse for me any pdf or text source for chatting has to be 100% efective and no one of this techs show that results and I am not feel conftable and profesional deploying this kind of solution with no 100% of success response. What do you think?

Nice! 👏

[@federico-bi2w]

Great! Thanks!

[@SKNti]

Thanks always

[@rahuldinesh2840]

The speed of retrieval and the Chinese characters doesn’t seem impressive.

[@ZukoBronja]

👍

[@loz6441]

This looks like an interesting topic, but simply reading slides is not a way to engage viewers!!!!! So I gave up after the first few slides and googled other sources.

[@thingX1x]

I discovered KAG accidentally haha 😂

[@mihaicarnuta]

very shallow explanation