- Stream Notes
- AI News
- Coding
- Let’s fix up our Vespa schema
- Reading up on Vespa blogs
- Billion-scale KNN
- Part 1
- They mention hundreds of billions of stars
- Star-sized vector datasets could happen with mixed data
- Trying to address trade-offs
- Accuracy of approximate nearest neighbor vs exact neighbor search
- Latency and scaling
- Scaling search volume (searches per second)
- Batch indexing vs real-time indexing and handling of updates and vector metadata
- Distributed search in large vector datasets which does not fit in a single node
- Cost, including total cost of ownership and dev efforts
- AI-Powered Vector Representations
- Pretrained Transformer Language Models for Search
- References BERT
- Feels old, but vocab used still applied today
- Zero-shot learning
- Multi-model representations
- Also need to convert Edamam calls to JSON to feed into Vespa
- Could use images to make vectors as well, since the images will have cuisine labels, titles, etc just like the ingredients and steps
- Follow along with https://blog.vespa.ai/text-image-search/
- TODO figure out prompting to make placeholder images of recipes in batches
- From Chat / Derail
- Offkai plans?
- #theunholydiver
- Take a look at Chatterino for chat; it has a lot of features for streamers and is faster than twitch itself
- #stream_kyle
- #dungspren
- if you’re trying to understand transformers and want a good article that dumbed it down
- Cooking question: is soaking veggies in water or vinegar to clean them/“remove the scum” a Japanese thing?
- #queuemina
- We raided yuzumiin
Socials
Related