AI GUIDEFrom a service to a system
In Base you used a finished application. At Advanced, separate the components so you can diagnose a result: the model, the instructions and context, any tools, and the application that assembles them. An improvement in one component does not prove that another changed.
Pretraining adjusts model parameters using a broad training process. Inference uses trained parameters to produce a response. Fine-tuning is additional training for an adaptation objective. Giving a model a document in a request is context, not evidence of fine-tuning. Retrieving that document from a collection changes what information reaches the request, not the model's parameters.
These distinctions matter when a team says “we need to train an AI on our company”. The actual need might be current document retrieval, consistent output fields or permission-aware actions. Each suggests a different intervention. Start by naming the observed failure, not the proposed technology.
Match the intervention to the failure
Fictional case: a support assistant gives yesterday's opening hours. First check whether it received today's source and whether the application selected the right location. Fine-tuning on old opening-hour examples would not solve a missing current lookup.
Second case: correct information repeatedly appears in an inconsistent report format. Try explicit field definitions, examples and structured output before considering training. Third case: a specialist classification task needs stable behaviour across many representative examples. Fine-tuning may become a candidate, but requires suitable data, evaluation and a comparison with a simpler baseline.
The choice is not always exclusive: an adapted model can also use retrieval and tools. Keep a record of which change you are testing. Otherwise, changing the model, prompt, source set and workflow at once makes it hard to explain a better or worse outcome.
Model selection is a comparison
Choose an evaluation set representative of your task, with an expected answer or rubric for each item. Include ordinary cases, ambiguous inputs and missing information. Keep the same input material and tool access across candidate models. Record model/version, settings, source versions, elapsed time and usage cost.
Measure the outcome you need: for document extraction, whether facts and unknowns are preserved; for code, whether relevant behaviour works; for translation, whether meaning and terminology survive. Do not substitute a general leaderboard for these checks.
Open model weights can give you more control over deployment and adaptation, but do not automatically mean no licence restrictions, no operating cost or better privacy. Hosted APIs trade some infrastructure work for provider dependence and data-handling choices. Check actual licences and service terms for the specific choice; this lesson makes no blanket licence claim.
Context, tokens and caching
Tokens are processing units used by language models; they are not a fixed one-to-one count of words. Context limits constrain what can fit into a request. A long context allowance does not make every included document relevant or resolve contradictory dates.
Some providers offer prompt caching for repeated input prefixes. That can reduce repeated processing under the provider's rules; it is not model training or a guarantee of persistent user memory. Separate cached input, application storage and a model's learned parameters when discussing retention or cost.
For a long company handbook, first ask whether every section is needed for each question. A focused relevant excerpt may be easier to check than a whole archive. If the answer needs information distributed across several sections, retrieval must preserve that coverage rather than choosing one superficially similar paragraph.
Design an experiment you can defend
Exercise: your fictional assistant extracts supplier, amount, currency and due date from ten invoice notes. Two notes omit dates and one contains contradictory totals. Write expected records before choosing a model. Missing dates must remain unknown and the contradiction must be flagged.
Compare a basic request with an explicit extraction contract using the same model and inputs. Then, if useful, compare another model with that same contract. Explain which change improved which case. A ten-case experiment is a learning baseline, not a population-level accuracy claim.
Completion criterion: you can identify whether the next intervention should target instructions, sources, a tool, model choice or training, and name the evidence that would change your decision. For implementation depth, follow the Hugging Face training chapter. Continue Advanced with A08, then AD02 on retrieved knowledge.
Sources
Read the original sources to explore the topic and check specific claims.