Give AI the right context and know where its knowledge comes from

Context is the material available for this response

A context window limits how much information a model can handle in a request. Conversation history, instructions and supplied material compete for that space; product behavior varies. Claude's context documentation explains the mechanism for one platform. A large limit does not guarantee that every detail will be used correctly.

In practice, provide the smallest sufficient packet: the goal, relevant source passages, constraints and desired output. Label document versions. If a conversation becomes confused, start a fresh task with a checked brief rather than assuming that adding more messages will repair it.

Four ways information can enter an answer

General model knowledge comes from training. Supplying a brief adds information for the current task. An application may retrieve relevant passages from a collection before asking the model to respond; this approach is often called retrieval-augmented generation, or RAG. Application memory may store selected preferences or facts for later use. These mechanisms are different.

RAG retrieves material and supplies it as context; it does not automatically retrain the model. Anthropic's retrieval explanation describes this pattern and the difficulty of retrieving the right passage. For a small task, pasting one permitted passage may be enough; a retrieval system becomes useful when the relevant material must be selected from a larger collection.

Saying “learn my English mistakes” is ambiguous. It could mean “use these examples in this chat,” “keep an external mistake log,” or “use an application's memory.” It does not by itself establish model fine-tuning, permanent storage or reliable recall. Keep a readable study record you control, then provide the current relevant part.

Fine-tuning is different again: it adapts model parameters through additional training. Giving examples in a prompt does not itself change those parameters. See Google's explanation of tuning and prompting. For a personal study record, begin with context and a readable log rather than assuming you need training infrastructure.

Worked example with a fictional policy

Give an assistant exactly these lines:

Source: DEMO club policy, version 2, fictional.

P1. Members may cancel a booked court up to 12 hours before the start.

P2. Later cancellations require a staff decision.

Task: Explain the cancellation deadline and whether a late cancellation is refundable.

Use this source only. Cite P1/P2. State what cannot be determined.

An acceptable editor-written answer is: “The cancellation deadline is 12 hours before the start (P1). A later cancellation needs a staff decision (P2). The passage does not state whether a refund is available.” A claim that the refund is automatic invents a policy. An answer citing P2 for a refund is still unsupported.

Now change P1 to 24 hours and label the new packet version 3. Ask the same question in a fresh conversation. The expected answer must use 24 hours and identify version 3. This is a simple check of source use, not proof that the system will handle every document correctly.

Retrieval has failure points

The wrong edition may be selected. A table may be read without its footnote. A relevant passage may not be found. A document can contain instructions that have nothing to do with your task. Treat source content as evidence to inspect, not as permission to change the task or reveal other information.

Keep source selection and answer verification separate. First ask, “Did we retrieve the right material?” Then ask, “Does it support this particular claim?” A correct-looking link answers neither question by itself. For a real policy decision, open the original and check its applicability.