Improving Model Output

Google's guide names the limitations explicitly, and the useful discipline is pairing each with the remedy that actually addresses it.

Limitation	What it looks like	The remedy that fits
Data dependency	Wrong or skewed on topics poorly represented in training	Grounding in your own data
Knowledge cutoff	Confident and out of date	Retrieval, or grounding in live search
Bias	Systematically different treatment across groups	Measurement by segment, curated data
Fabrication	Fluent, confident, invented	Grounding, citations, an explicit way to say "I don't know"
Edge cases	Fails on the unusual input nobody tested	Human in the loop where consequence is high

The mistake this table is designed to prevent: reaching for fine-tuning as a general-purpose fix. Fine-tuning shapes behavior, tone, and format. It does not install knowledge and it does not fix fabrication. If the problem is "it doesn't know our stuff", the answer is grounding.

Human in the loop is the remedy for consequence rather than for accuracy. It does not make the system better; it makes the failures catchable. Place it where being wrong is expensive.

Keeping it honest after launch needs four things, and this is the part that gets skipped: version what you deploy, track performance against a fixed set of cases, monitor for drift as real inputs move away from what you tested, and set a review cadence with a named owner. A system nobody is watching does not stay level. It degrades quietly while everything appears to work.

The core techniques, in the order you should reach for them:

• Zero-shot: just ask. Fine for common tasks with obvious output shape.
• One-shot and few-shot: supply one or a few worked examples. The fix when the format is wrong.
• Role prompting: tell the model who it is writing as and for whom. Cheap, and it moves register a long way.
• Prompt chaining: break a complex task into steps, feeding each output into the next. The fix when one prompt is trying to do too much.

The advanced pair and when they earn their cost:

• Chain-of-thought: ask for explicit reasoning before the answer. Use when the format is right but the thinking is wrong. It costs tokens and latency, so it is not a default.
• ReAct: a loop where the model reasons, takes an action with a tool, observes the result, and repeats. This is what underlies agents. Use when the task genuinely requires gathering information the model does not start with.

The sampling controls are worth understanding because people tune them superstitiously:

• Output length: a cap. Set it deliberately; truncated answers are a common and avoidable failure.
• Temperature: how much randomness in word choice. Low for factual and consistent work, higher for creative variety. Low temperature is not a fix for fabrication. It makes wrong answers more consistent, not more correct.
• Top-p (nucleus sampling): an alternative way to constrain word choice. Adjust one or the other, not both.
• Safety settings: thresholds for filtering harmful content, and a real governance decision rather than a technical default.

The temperature misconception is the one to carry into the exam: turning it down makes output more predictable, not more accurate.

Grounding means making the model answer from supplied material rather than from what it absorbed in training. Google's guide splits the sources three ways, and the distinction has real governance consequences:

• First-party enterprise data: your own documents, records, and systems. Most valuable, and it carries every permission and privacy obligation you already have.
• Third-party data: licensed external sources. Check what the license actually permits for AI use; this is a live legal question.
• World data: public web information via search grounding. Best for current events and general facts, and least controllable in provenance.

Retrieval-augmented generation is the mechanism for the first two: find the passages relevant to the question, attach them to the prompt, answer from them. The effect on output is the thing to be able to state plainly. It moves the failure mode from invented to possibly misread, which is a large improvement because misreadings are catchable and inventions are not.

The managed options span prebuilt retrieval you point at your content, retrieval APIs for building your own pipeline, and grounding directly in web search results. The build-versus-buy judgment here favors managed more strongly than people expect, because retrieval quality is where these systems actually fail and it is unglamorous work to get right.

One diagnostic habit worth taking away: when a grounded system starts producing confident nonsense, look at what was retrieved before you touch the prompt. If the wrong material came back, the model answered faithfully from bad input and no prompt change will fix it.