13 posts found
Tokens, not words: how a model reads your text
Models do not see characters or words. They see tokens — and once you know how text becomes tokens, several odd behaviou...
Embeddings explained without the linear algebra
An embedding turns text into coordinates, and nearby coordinates mean related meaning. That single idea is what makes se...
Training, fine-tuning and prompting are three different tools
Teams reach for fine-tuning when they need context, and for prompting when they need behaviour. Knowing which problem ea...
The context window is a budget, not a memory
Bigger context windows did not give models memory. They gave you a larger envelope to fill on every single request — and...
Choosing between a large model and a small one
Most production traffic does not need the largest model available. Routing by task rather than defaulting to the top of...
Why agents fail on long tasks
An agent that handles five steps beautifully can fall apart at twenty. The reason is rarely reasoning — it is that the c...
Giving an agent memory without giving it amnesia
Memory is not a feature you enable. It is a set of decisions about what to keep, what to summarise and what to let go —...
RAG in one page: retrieve, rank, answer
Retrieval-augmented generation is three steps and a lot of tuning. The architecture is simple; the quality lives almost...
Chunking is the part of RAG nobody tunes
Teams spend weeks on rerankers and leave chunking at the default. It is usually the other way round that pays — how you...
Prompt caching: the cheapest speedup you are not using
If every request begins with the same two thousand tokens of instructions, you are paying full price to resend them. Ord...
Rate limits, retries and the backoff you actually need
Every AI feature meets a 429 eventually. Whether that is a blip or an outage depends on retry logic written before you n...
The cost model of an AI feature
Per-token pricing looks trivial until you multiply by retries, conversation history and the context you resend on every...