Notes from a journey of compounding curiosity.
When building sophisticated agents that have more than a handful of tools to call, I've often found the inbuilt structured output/json tool calling methods provided by LMA APIs come up short. Intuitively, one of the reasons for this is that when the structured output is enabled, there is no room for chain of thought, text or inbuilt support for comments amongst the JSON output of tool calls. In…
When building a RAG pipeline you'll probably reach for a vector store to store embeddings of document chunks, which are then retrieved and put into context at query time. This works well if your users are asking single fact queries where the answer can be found in a relevant document chunk. But if your users want to ask more complex questions where the answer requires information spread across the…
Impressive new language models like Llama and Mistral have broadened the accessibility of AI training. If you want to fine-tune a model with your own data, it's now relatively easy to do with tools like Axolotl and a few dollars spent on a GPU cloud. But if you want to go deeper and train larger models or try new methods, the cloud bill can quickly rack up. Renting 8 A100's on AWS will set you…