ahnaf @ahnafoi - Twitter Profile

over 1 year ago

🤯 FAST KG-based Data Retrieval RAG After some tweaks and a hybrid retrieval/graph entity creation pipeline, my system processed 13 queries semantic queries in just 10.0 seconds (±0.5-1s) in a small test set filled with synthetic data That's performing really well for this type of task! Watch the test run demo Anyone else tackling speed in KG-enhanced LLMs? Would love to hear your approaches. Maybe we can collaborate and build a benchmarking suite. #KG #LLM #AI #NLP #RetrievalAugmentation #KnowledgeGraph #Tech #MachineLearning #RAG

0

34

ahnaf @ahnafoi

over 1 year ago

I noticed LLM chat interfaces could benefit from better persistent memory, so I developed my own graph-based semantic memory system. It stores and traverses data based on a KG architecture with less rigidity on schema It supports live updates and self-reasoning over connections. It uses embeddings to aid itself, the LLM actively navigates the graph using iterative top-k searches to find deeply embedded knowledge, this is demonstrated in the first intentional fail query result in the video below A small demo of an edge case to test potential hallucination after previously inputting data which may generally cause hallucinations: https://t.co/t8ovaj38KK I’d love to be able to compare this to the memory implementation of ChatGPT @OpenAI @sama and benchmark it across volume and multi hop, but the credits needed to generate the synthetic data is going to leave a dent in my bank account which I can’t justify as a student (let alone one getting the method correct in my first try) Any help with benchmarking efforts would be greatly appreciated, I do have a few ideas but I have no clue whether they are good or bad, and this is assuming I can afford it @karpathy @langchain @goodfellow_ian @OfficialLoganK @hwchase17 My DMs are open if anyone is interested in collaborating or helping in coming up with a frame work to test this system

ahnafoi's tweet photo. I noticed LLM chat interfaces could benefit from better persistent memory, so I developed my own graph-based semantic memory system. It stores and traverses data based on a KG architecture with less rigidity on schema

It supports live updates and self-reasoning over connections. It uses embeddings to aid itself, the LLM actively navigates the graph using iterative top-k searches to find deeply embedded knowledge, this is demonstrated in the first intentional fail query result in the video below

A small demo of an edge case to test potential hallucination after previously inputting data which may generally cause hallucinations:
https://t.co/t8ovaj38KK

I’d love to be able to compare this to the memory implementation of ChatGPT @OpenAI @sama and benchmark it across volume and multi hop, but the credits needed to generate the synthetic data is going to leave a dent in my bank account which I can’t justify as a student (let alone one getting the method correct in my first try)

Any help with benchmarking efforts would be greatly appreciated, I do have a few ideas but I have no clue whether they are good or bad, and this is assuming I can afford it

@karpathy @langchain @goodfellow_ian @OfficialLoganK @hwchase17

My DMs are open if anyone is interested in collaborating or helping in coming up with a frame work to test this system

0

56

ahnafoi retweeted

d@x @groktheworm

over 3 years ago

I'm speechless. Not peer-reviewed yet but a submitted paper. The 'presented images' were shown to a group of humans. The 'reconstructed images' were the result of an fMRI output to Stable Diffusion. In other words, #stablediffusion literally read people's minds. Source 👇