Thank you! I'm definitely planning to make more content with open source LLMs, like how to fine tune and deploy them to an API you can use just like you would use the OpenAI or Azure APIs 🙂
The main pointer is to make the graph objects as easily distinguishable as possible, i.e name your node-types, relationship-types and their attributes in a way where the names are descriptive and do not get easily mixed. Really this applies for schema design in general, but is especially helpful to reduce the chances that the LLM will mix any of the objects when generating Cypher. Good luck!
Hey, sure! Performance generating the cypher will of course vary with open-source models but otherwise the principles are exactly same (: You might also consider running OpenAI's models on Azure if privacy is a concern. That way you will have Azure's enterprise-grade data privacy guarantees, and you can also choose the region where the model is hosted
brah what happens when u use random stream of text. imagine u have a k. graph already, and u continiously stream new data to it. how do u make sure that the entities stay the same, if in round 1 entity was John Mayer, and in round 2, the text used John M. or Johnny Mayer how can u sync this up?
Hey, great question. This problem of "syncing" similar entities is a key challenge in making this work in real-life. There are some promising methods, which you can learn more about in these two links: margin.re/2023/06/entity-resolution-in-reagent/ (in the second part) ua-cam.com/video/dNGV4sLkOcA/v-deo.html I'm not sure how soon I'll manage it, but I definitely want to make a video on this as well, hopefully before too long
same here I get the following error: ValidationError: 3 validation errors for GraphCypherQAChain cypher_generation_chain field required (type=value_error.missing) qa_chain field required (type=value_error.missing) graph_schema field required (type=value_error.missing would also like to know how to solve this issue
Problem is the code isn't parsed properly and any error will end up showing 'cypher_query' not defined. Mine was OPENAI_API_BASE in env without value that i left there and made the whole thing bug because it was overwritting the openai API URL in my from langchain_community.chat_models import ChatOpenAI
Actually wasn't aware of SurrealDB before now, looks interesting! I think it should work with the same principles, but with the caveat that SurrealQL being a new dialect of SQL, I think LLMs will struggle to generate it reliably. This is already a significant challenge for Cypher (as I mention in the video) and even standard SQL, so for SurrealQL I think you'd definitely need to fine-tune a model for it.
Thanks for the feedback! I'm still figuring out how to best edit these videos, and here I tried to balance it by only having music in the intro and outro. Do you think the music is too loud, or would you prefer no music at all?
@johannesjolkkonen I thought the reason was to avoid your audio being scrapped by skynet! I personally cannot understand well when there is any music on the background and is distracting, but others might disagree. I only watch channels without music as I want the hear clearly.
Thanks for sharing your experience and knowledge on classic coding problems beginners run into!
An absolute treat! Thank you again!
Awesome tutorial would love to see "How to Build Knowledge Graphs With open source LLMs "
Thank you!
I'm definitely planning to make more content with open source LLMs, like how to fine tune and deploy them to an API you can use just like you would use the OpenAI or Azure APIs 🙂
@@johannesjolkkonen Awesome 😍
waiting for this
@@johannesjolkkonen thanks, looking forward to building custom Knowledge Graphs from our own data using open source llm.
I'm gung ho for Open Source❗ What about Neo4J -- can you replace it too❓
Really enjoy your tutorials! We should once do a session together, will be fun!
Hey Dmitry, glad to hear it!
Let's make it happen. Sent you a connect-invite on LinkedIn 🙂
Love the solution. I would really appreciate a larger font size though.
Thank you for sharing.
Challenge is coming up with domain specific schema when dealing with unstructured data i.e. mostly text. Any pointers about that ?
The main pointer is to make the graph objects as easily distinguishable as possible, i.e name your node-types, relationship-types and their attributes in a way where the names are descriptive and do not get easily mixed.
Really this applies for schema design in general, but is especially helpful to reduce the chances that the LLM will mix any of the objects when generating Cypher. Good luck!
Good but can you do these with just ChatOpenAI not Microsoft stuff Azure etc ? that way its more open source?
We use all of these resources except we are not allowed to send anything to openai can this be accomplished with a local llm?
Hey, sure! Performance generating the cypher will of course vary with open-source models but otherwise the principles are exactly same (:
You might also consider running OpenAI's models on Azure if privacy is a concern. That way you will have Azure's enterprise-grade data privacy guarantees, and you can also choose the region where the model is hosted
brah what happens when u use random stream of text. imagine u have a k. graph already, and u continiously stream new data to it. how do u make sure that the entities stay the same, if in round 1 entity was John Mayer, and in round 2, the text used John M. or Johnny Mayer how can u sync this up?
Hey, great question. This problem of "syncing" similar entities is a key challenge in making this work in real-life. There are some promising methods, which you can learn more about in these two links:
margin.re/2023/06/entity-resolution-in-reagent/ (in the second part)
ua-cam.com/video/dNGV4sLkOcA/v-deo.html
I'm not sure how soon I'll manage it, but I definitely want to make a video on this as well, hopefully before too long
Is there any way to apply this to the HTML Tree?
NameError: name 'cypher_query' is not defined
same here I get the following error:
ValidationError: 3 validation errors for GraphCypherQAChain cypher_generation_chain field required (type=value_error.missing) qa_chain field required (type=value_error.missing) graph_schema field required (type=value_error.missing
would also like to know how to solve this issue
Late to the party, but it's because of your connection string to your neo4j database. Check that you have configured it correctly.
Problem is the code isn't parsed properly and any error will end up showing 'cypher_query' not defined.
Mine was OPENAI_API_BASE in env without value that i left there and made the whole thing bug because it was overwritting the openai API URL in my from langchain_community.chat_models import ChatOpenAI
is SurrealDB any good for these ?
Actually wasn't aware of SurrealDB before now, looks interesting!
I think it should work with the same principles, but with the caveat that SurrealQL being a new dialect of SQL, I think LLMs will struggle to generate it reliably.
This is already a significant challenge for Cypher (as I mention in the video) and even standard SQL, so for SurrealQL I think you'd definitely need to fine-tune a model for it.
I have to really thank you ، how i can contact you
You look like Homelander. lol
a shame is has music
Thanks for the feedback! I'm still figuring out how to best edit these videos, and here I tried to balance it by only having music in the intro and outro.
Do you think the music is too loud, or would you prefer no music at all?
@johannesjolkkonen I thought the reason was to avoid your audio being scrapped by skynet! I personally cannot understand well when there is any music on the background and is distracting, but others might disagree. I only watch channels without music as I want the hear clearly.
Second for no music 🙏🏼
@@johannesjolkkonen I'm not here for no stinking music❗ 😏 I need info as efficient as possible - No fluff. Good Job overall 👍