Great video. I think if you focus on using open source tech or to be blunt making this as cheap as possible you'll track in more viewers! :) Thank you for the great video. I'm a TPM and have to plan a graphRAG for tens of thousands of video transcripts and cost matters a lot. Still learning if it'll even be possible!
@3:18 "Which I experienced" 🫠Man, I feel your pain. I uploaded 4 classes transcripts to a default everything Bedrock/OpenSearch and it was 5$ a day! Imagine 50000 classes... The name of the game here is keeping costs down.
That is a problem now - things get cheaper. Also a question whether using highly expensive models is best here.... Claude i.e. is significantly cheaper. Anyone thinks that the LLM is the problem - it is not, the memory is.
This is amazing sir. Can you also post a video on how to do this via groq. I couldn't figure out the endpoints of groq. Since, there is not much utilization of the embedding models I am fine using the openai models for that but for the main model itself.
Great job Mervin, but you said in the last video that you will do it with Ollama. Please keep an equal focus on OpenAI models and Ollama-based ones as well. (Because many TH-camrs are not focusing on OpenSource models - in YT views point of view as well).
It would help to create in detail a video that reviews each of the internal text prompts that exist in graphrag. Even if one is not using graphrag, it should be educational to be aware of what the prompts are trying to accomplish and why.
Thank you for the great video. Is there a maximum input size for creating a GraphRAG? Every time I try to insert my documents, I get an error with "create_summarized_entities."
Hi Mervin, I may ask a stupid question here, I tried the GraphRAG, it is really good! But the whole system seems a BlackBox to me, I can not control the way how to extract the entity, relation, create the graph and community. Do you think it is possible to do this level modification for any customization purpose? If so, could you show the way to do so? Thank you！
hello sir can we add new data to the existing knowledge graph ? if so how to do it ? Like a user adds a new file in chatbot do we need to append the input folder with new txt or do the entire process again ?
Great video. But I’m trying to run this locally. I got it to work with qwen2:7b but can’t get it to do the embedding with mimic text embedding. It fails running the final part of indexing. Running this off my local ollama server. It will be slow for indexing but I could load up a bunch of docs and let it run. I’d like to see a openwebui pipeline for this.
good topic but it doesn't seem to provide enough cost reduction to make it production ready. Couldn't we keep the cost down significantly if we fine-tuned a custom LLM and make private API calls to it ? with 1 LLM you can potentially serve hundreds of clients.
This is pretty frustrating - it's not addressing any of the concerns about contextual hallucinations and how to spot them. No evidence at all that this is a step change method from that perspective, it's like a marginal increase over standard RAG that gives you refs from the contexts. But is the output actually correct?
Great video. I think if you focus on using open source tech or to be blunt making this as cheap as possible you'll track in more viewers! :) Thank you for the great video. I'm a TPM and have to plan a graphRAG for tens of thousands of video transcripts and cost matters a lot. Still learning if it'll even be possible!
Great job Mervin. Please publish more videos about GraphRAG
@3:18 "Which I experienced" 🫠Man, I feel your pain. I uploaded 4 classes transcripts to a default everything Bedrock/OpenSearch and it was 5$ a day! Imagine 50000 classes... The name of the game here is keeping costs down.
That is a problem now - things get cheaper. Also a question whether using highly expensive models is best here.... Claude i.e. is significantly cheaper. Anyone thinks that the LLM is the problem - it is not, the memory is.
This is amazing sir. Can you also post a video on how to do this via groq. I couldn't figure out the endpoints of groq. Since, there is not much utilization of the embedding models I am fine using the openai models for that but for the main model itself.
Thanks, excellent video showing how to interface graphrag with Python code👍
Awesome tutorial! 👏
Thanks Mervin for the amazing video!
Does anyone know if a Neo4j backend is possible with Graphrag?
Great video. Thanks!
Just as a note, the link to your code apparently does not include the Chainlit powered UI version
Great job Mervin, but you said in the last video that you will do it with Ollama. Please keep an equal focus on OpenAI models and Ollama-based ones as well. (Because many TH-camrs are not focusing on OpenSource models - in YT views point of view as well).
Love the explanation, can you also do a video maybe on the Atomic Agents AI library? It's extremely elegant IMO
github URL: github.com/KennyVaneetvelde/atomic_agents
It would help to create in detail a video that reviews each of the internal text prompts that exist in graphrag. Even if one is not using graphrag, it should be educational to be aware of what the prompts are trying to accomplish and why.
does this support Ollama and LLMs other than GPT ?
If it's the Microsoft one, it does support Ollama
0:50 he describes doing that
I got ollama to work with qwen2:7b but can’t get embedding to work with nomic at the moment.
It supports Groq API??
Hello Mervin, for some reason the Inputs folder is not being created now.
Thank you for the great video. Is there a maximum input size for creating a GraphRAG? Every time I try to insert my documents, I get an error with "create_summarized_entities."
Hi Mervin, I may ask a stupid question here, I tried the GraphRAG, it is really good! But the whole system seems a BlackBox to me, I can not control the way how to extract the entity, relation, create the graph and community. Do you think it is possible to do this level modification for any customization purpose? If so, could you show the way to do so? Thank you！
hello sir can we add new data to the existing knowledge graph ? if so how to do it ? Like a user adds a new file in chatbot do we need to append the input folder with new txt or do the entire process again ?
Great video. But I’m trying to run this locally. I got it to work with qwen2:7b but can’t get it to do the embedding with mimic text embedding. It fails running the final part of indexing. Running this off my local ollama server. It will be slow for indexing but I could load up a bunch of docs and let it run. I’d like to see a openwebui pipeline for this.
good topic but it doesn't seem to provide enough cost reduction to make it production ready. Couldn't we keep the cost down significantly if we fine-tuned a custom LLM and make private API calls to it ? with 1 LLM you can potentially serve hundreds of clients.
can this run fully locally using ollama?
Is it possible to use it with Claude?
Ollama FIRST!
Would I be able to use Vllm to reduce costs for this?
I got ollama with qwen2:7b to work but can’t get the local embedding to work yet
is it possible to use gemini api?
This is pretty frustrating - it's not addressing any of the concerns about contextual hallucinations and how to spot them. No evidence at all that this is a step change method from that perspective, it's like a marginal increase over standard RAG that gives you refs from the contexts. But is the output actually correct?
Just use your own free hugging face agent and you are done in 10 minutes flat 😊
This is not amazing. This is throwing money to API services.
care to elaborate?
@@awakenwithoutcoffee Why should I explain the obvious? It was fun to see the local variant (at least the search) from the next video.