Tech & Future
Understand the machines rewriting everything
Novice
Subdomain
How to get useful, reliable output from today's models. Most quality issues are prompt issues, not model issues.
Hallucination
Confident, fluent, totally wrong. The signature failure of LLMs.
Open
Prompting
How you ask is most of what you'll get.
Open
Retrieval-Augmented Generation (RAG)
RAG is a pattern where, before the model answers, relevant chunks of your own documents are fetched and pasted into the prompt, so the answer is grounded in your sources instead of the model's frozen memory.
Open
Chain-of-thought prompting
A prompting technique that asks the model to write out intermediate reasoning steps before giving a final answer, which raises accuracy on multi-step problems.
Open
Vector databases
Specialised databases that index and search dense vector embeddings using approximate nearest neighbour algorithms for semantic similarity queries.
Open
LLM evaluation metrics
Quantitative measures used to score language model outputs, ranging from text overlap metrics like BLEU and ROUGE to perplexity and benchmark suites like MMLU.
Open
Parameter
A tunable weight inside a model adjusted during training.
RAG
Retrieval-Augmented Generation, giving a model external documents to ground its answers.
Overfitting
Memorising training data so well the model fails on new inputs.
Inference
Using a trained model to produce outputs.
Fine-tuning
Continuing to train a model on a smaller, more specific dataset to specialise it.
Embedding
A vector of numbers representing the meaning/features of a chunk of input.
RLHF
Reinforcement Learning from Human Feedback, alignment via human ratings.
Hallucination
A fluent, confident output that's factually wrong.
Token
The smallest input/output unit an LLM works with, typically 3 to 4 characters in English.
Context window
The maximum tokens a model can consider in a single forward pass.
Vector database
A store optimised for nearest-neighbour search over embeddings.
MoE
Mixture of Experts. Routes tokens to specialised sub-networks so active parameters per token stay small.